Single-pass class method indexing in pyrefly coverage

[jorenham]

Summary

The parse_classes function was re-iterating every binding key in the module for each class, just to find the methods of that class. This changes the code (not the behavior) so that this now only requires a single pass.

Here's a quick little benchmark:

Corpus	before	after	speedup
synthetic, 1 module, 2,000 classes	0.457s	0.274s	1.67x
synthetic, 1 module, 6,000 classes	4.182s	0.803s	5.2x
scipy-stubs, 589 modules (and 1 maintainer 😅 )	1.766s	1.783s	~1.0x

So it looks a bit like this could be a quadratic speedup (in the #classes/module sense)

Test Plan

Tests still pass

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[NathanTempest]

Thanks @jorenham for your continued efforts in refactoring and improving pyrefly coverage !!

Replacing the per-class O(classes × bindings) rescan with a single grouping pass is the right fix, and the benchmark makes the improvement clear. Importing and merging.

[meta-codesync]

@NathanTempest has imported this pull request. If you are a Meta employee, you can view this in D108442250.

[yangdanny97]

Review automatically exported from Phabricator review in Meta.

[meta-codesync]

@NathanTempest merged this pull request in e5e8e89.