Instead of locking and unlocking the mutex for every symbol,
we gather what we can and only lock the mutex after processing
a module.
Before:
```
$ ./target/profiling/ty_completion_bench ~/astral/relatedclones/scratch-home-assistant/homeassistant/scratch.py 1 -q --iters 30
total elapsed for initial completions request: 579.057075ms
total elapsed: 5.627666455s, time per completion request: 187.588881ms
```
After:
```
$ ./target/profiling/ty_completion_bench ~/astral/relatedclones/scratch-home-assistant/homeassistant/scratch.py 1 -q --iters 30
total elapsed for initial completions request: 565.911487ms
total elapsed: 5.254306446s, time per completion request: 175.143548ms
```
This is a very naive/simplistic "improvement." We can probably do
better.
1f6a87f770