Changes the red-knot benchmark to run on the stdlib "tomllib" library
(which is self-contained, four files, uses type annotations) instead of
on very small bits of handwritten code.
Also remove the `without_parse` benchmark: now that we are running on
real code that uses typeshed, we'd either have to pre-parse all of
typeshed (slow) or find some way to determine which typeshed modules
will be used by the benchmark (not feasible with reasonable complexity.)
## Test Plan
`cargo bench -p ruff_benchmark --bench red_knot`
