Optimise venv identification for mid/large-sized directory trees

[alexander-beedie]

Summary

Overview

• When pyrefly check runs over a package without a configured interpreter, it auto-discovers a project's venv.

• If "pyvenv.cfg" isn't immediately found in the project root or ".venv" / "venv", it falls back to an unfiltered WalkDir (to depth 3) over the entire project, testing every filename against the interpreter regex.

• For large projects, this can be a significant fixed cost. For example, profiling showed it taking up ~20% of the total runtime for checking homeassistant, as that package has ~2,700 directories (!!). For small projects there is no change.

Solution

• walk_interpreter gains a simple "should_descend" predicate that prunes a directory's subtree when the predicate rejects it; a venv-aware pruning rule is supplied in find_in_dir.

• The prune is a strict superset of the reachable interpreters, so which interpreter the find resolves is unchanged, but the non-venv source fan-out is skipped.

• In the extreme homeassistant case, this reduces ~2,700 opendir calls down to only ~15, making the check essentially free (instead of ~20% of the runtime) 🚀

Test Plan

All existing tests pass, and two new venv discovery tests were added (which actually represent the majority of this PR as the code change itself is small).

Benchmarks¹

Timed across a sample of the mypy_primer projects. Speedup is proportional to directory-tree size; significant on large trees, negligible on small flat projects.

dirs ≤ 3 = how many directories (with depth ≤ 3) in the project.
modules = how many modules in the project.
before = pyrefly on main branch.
after = main + this PR.

project	dirs ≤ 3	modules	before	after	Δ%
homeassistant	2,736	9,811	2.010	1.611	-19.8%
poetry	124	191	0.172	0.151	-12.6%
sphinx	360	243	0.253	0.229	-9.4%
pylint	129	51	0.178	0.162	-9.0%
hypothesis	66	103	0.172	0.157	-8.7%
sympy	153	1,560	0.358	0.334	-6.6%
pydantic	60	104	0.170	0.164	-3.4%
pandas	129	1,459	0.407	0.396	-2.6%
mypy	61	1,098	0.394	0.384	-2.5%
scipy	137	925	0.429	0.425	-1.0%

Footnotes

1. Test machine: Apple Silicon M3 Max. ↩