[red-knot] Trust module-level undeclared symbols in stubs (#17577)

## Summary

Many symbols in typeshed are defined without being declared. For
example:
```pyi
# builtins:
IOError = OSError

# types
LambdaType = FunctionType
NotImplementedType = _NotImplementedType

# typing
Text = str

# random
uniform = _inst.uniform

# optparse
make_option = Option

# all over the place:
_T = TypeVar("_T")
```

Here, we introduce a change that skips widening the public type of these
symbols (by unioning with `Unknown`).

fixes #17032

## Ecosystem analysis

This is difficult to analyze in detail, but I went over most changes and
it looks very favorable to me overall. The diff on the overall numbers
is:
```
errors: 1287 -> 859 (reduction by 428)
warnings: 45 -> 59 (increase by 14)
```

### Removed false positives

`invalid-base` examples:

```diff
- error[lint:invalid-base] /tmp/mypy_primer/projects/pip/src/pip/_vendor/rich/console.py:548:27: Invalid class base with type `Unknown | Literal[_local]` (all bases must be a class, `Any`, `Unknown` or `Todo`)
- error[lint:invalid-base] /tmp/mypy_primer/projects/tornado/tornado/iostream.py:84:25: Invalid class base with type `Unknown | Literal[OSError]` (all bases must be a class, `Any`, `Unknown` or `Todo`)
- error[lint:invalid-base] /tmp/mypy_primer/projects/mitmproxy/test/conftest.py:35:40: Invalid class base with type `Unknown | Literal[_UnixDefaultEventLoopPolicy]` (all bases must be a class, `Any`, `Unknown` or `Todo`)
```

`invalid-exception-caught` examples:

```diff
- error[lint:invalid-exception-caught] /tmp/mypy_primer/projects/cloud-init/cloudinit/cmd/status.py:334:16: Cannot catch object of type `Literal[ProcessExecutionError]` in an exception handler (must be a `BaseException` subclass or a tuple of `BaseException` subclasses)
- error[lint:invalid-exception-caught] /tmp/mypy_primer/projects/jinja/src/jinja2/loaders.py:537:16: Cannot catch object of type `Literal[TemplateNotFound]` in an exception handler (must be a `BaseException` subclass or a tuple of `BaseException` subclasses)
```

`unresolved-reference` examples


7a0265d36e/cloudinit/handlers/jinja_template.py (L120-L123)
(we now understand the `isinstance` narrowing)

```diff
- error[lint:unresolved-attribute] /tmp/mypy_primer/projects/cloud-init/cloudinit/handlers/jinja_template.py:123:16: Type `Exception` has no attribute `errno`
```

`unknown-argument` examples


https://github.com/hauntsaninja/boostedblob/blob/master/boostedblob/request.py#L53

```diff
- error[lint:unknown-argument] /tmp/mypy_primer/projects/boostedblob/boostedblob/request.py:53:17: Argument `connect` does not match any known parameter of bound method `__init__`
```

`unknown-argument`

There are a lot of `__init__`-related changes because we now understand
[`@attr.s`](3d42a6978a/src/attr/__init__.pyi (L387))
as a `@dataclass_transform` annotated symbol. For example:

```diff
- error[lint:unknown-argument] /tmp/mypy_primer/projects/attrs/tests/test_hooks.py:72:18: Argument `x` does not match any known parameter of bound method `__init__`
```

### New false positives

This can happen if a symbol that previously was inferred as `X |
Unknown` was assigned-to, but we don't yet understand the assignability
to `X`:


https://github.com/strawberry-graphql/strawberry/blob/main/strawberry/exceptions/handler.py#L90

```diff
+ error[lint:invalid-assignment] /tmp/mypy_primer/projects/strawberry/strawberry/exceptions/handler.py:90:9: Object of type `def strawberry_threading_exception_handler(args: tuple[type[BaseException], BaseException | None, TracebackType | None, Thread | None]) -> None` is not assignable to attribute `excepthook` of type `(_ExceptHookArgs, /) -> Any`
```

### New true positives


6bbb5519fe/tests/tracer/test_span.py (L714)

```diff
+ error[lint:invalid-argument-type] /tmp/mypy_primer/projects/dd-trace-py/tests/tracer/test_span.py:714:33: Argument to this function is incorrect: Expected `str`, found `Literal[b"\xf0\x9f\xa4\x94"]`
```

### Changed diagnostics

A lot of changed diagnostics because we now show `@Todo(Support for
`typing.TypeVar` instances in type expressions)` instead of `Unknown`
for all kinds of symbols that used a `_T = TypeVar("_T")` as a type. One
prominent example is the `list.__getitem__` method:

`builtins.pyi`:
```pyi
_T = TypeVar("_T")  # previously `TypeVar | Unknown`, now just `TypeVar`

# …

class list(MutableSequence[_T]):
    # …
    @overload
    def __getitem__(self, i: SupportsIndex, /) -> _T: ...
    # …
```

which causes this change in diagnostics:
```py
xs = [1, 2]
reveal_type(xs[0])  # previously `Unknown`, now `@Todo(Support for `typing.TypeVar` instances in type expressions)`
```

## Test Plan

Updated Markdown tests

crates/red_knot_python_semantic/resources/mdtest/scopes/eager.md

@@ -404,7 +404,7 @@ x = int
 class C:
     var: ClassVar[x]
 
-reveal_type(C.var)  # revealed: Unknown | str
+reveal_type(C.var)  # revealed: str
 
 x = str
 ```

crates/red_knot_python_semantic/resources/mdtest/subscript/lists.md

@@ -12,7 +12,7 @@ x = [1, 2, 3]
 reveal_type(x)  # revealed: list
 
 # TODO reveal int
-reveal_type(x[0])  # revealed: Unknown
+reveal_type(x[0])  # revealed: @Todo(Support for `typing.TypeVar` instances in type expressions)
 
 # TODO reveal list
 reveal_type(x[0:1])  # revealed: @Todo(specialized non-generic class)

crates/red_knot_python_semantic/resources/mdtest/type_properties/is_singleton.md

@@ -128,10 +128,7 @@ python-version = "3.10"
 import types
 from knot_extensions import static_assert, is_singleton
 
-# TODO: types.NotImplementedType is a TypeAlias of builtins._NotImplementedType
-# Once TypeAlias support is added, it should satisfy `is_singleton`
-reveal_type(types.NotImplementedType)  # revealed: Unknown | Literal[_NotImplementedType]
-static_assert(not is_singleton(types.NotImplementedType))
+static_assert(is_singleton(types.NotImplementedType))
 ```
 
 ### Callables

crates/red_knot_python_semantic/src/semantic_index/symbol.rs

@@ -115,6 +115,10 @@ impl<'db> ScopeId<'db> {
         self.node(db).scope_kind().is_function_like()
     }
 
+    pub(crate) fn is_module_scope(self, db: &'db dyn Db) -> bool {
+        self.node(db).scope_kind().is_module()
+    }
+
     pub(crate) fn is_type_parameter(self, db: &'db dyn Db) -> bool {
         self.node(db).scope_kind().is_type_parameter()
     }
@@ -263,6 +267,10 @@ impl ScopeKind {
         matches!(self, ScopeKind::Class)
     }
 
+    pub(crate) fn is_module(self) -> bool {
+        matches!(self, ScopeKind::Module)
+    }
+
     pub(crate) fn is_type_parameter(self) -> bool {
         matches!(self, ScopeKind::Annotation | ScopeKind::TypeAlias)
     }

crates/red_knot_python_semantic/src/symbol.rs

@@ -593,9 +593,19 @@ fn symbol_by_id<'db>(
                 "__slots__" | "TYPE_CHECKING"
             );
 
+            if scope.is_module_scope(db) && scope.file(db).is_stub(db.upcast()) {
+                // We generally trust module-level undeclared symbols in stubs and do not union
+                // with `Unknown`. If we don't do this, simple aliases like `IOError = OSError` in
+                // stubs would result in `IOError` being a union of `OSError` and `Unknown`, which
+                // leads to all sorts of downstream problems. Similarly, type variables are often
+                // defined as `_T = TypeVar("_T")`, without being declared.
+
+                inferred.into()
+            } else {
                 widen_type_for_undeclared_public_symbol(db, inferred, is_considered_non_modifiable)
                     .into()
             }
+        }
         // Symbol has conflicting declared types
         Err((declared, _)) => {
             // Intentionally ignore conflicting declared types; that's not our problem,

crates/red_knot_python_semantic/src/types/signatures.rs

@@ -1631,10 +1631,7 @@ mod tests {
         assert_eq!(a_name, "a");
         assert_eq!(b_name, "b");
         // Parameter resolution deferred; we should see B
-        assert_eq!(
-            a_annotated_ty.unwrap().display(&db).to_string(),
-            "Unknown | B"
-        );
+        assert_eq!(a_annotated_ty.unwrap().display(&db).to_string(), "B");
         assert_eq!(b_annotated_ty.unwrap().display(&db).to_string(), "T");
     }