Cache pkgutil namespace probes during import resolution

[alexander-beedie]

Summary

Resolving a module and its submodules walks each path component, and every intermediate package's __init__.py is probed with is_pkgutil_namespace.

Crucially the same __init__ files are re-opened/re-read once per resolved submodule, leading to a lot of redundant calls (each of which is a File::open + 4KiB read + a regex to check for pkgutil.extend_path).

Solution

Memoise the result in a new pkgutil_cache on DirEntryCache, keyed by __init__ path, sharing the same "per-loader, never-invalidated" lifetime as the existing entry cache. The existing cache name was updated to entry_cache to help differentiate/clarify now there are two caches.

Impact

Reduces ~14,900 reads to ~1,150 when testing against the core package.
(One per unique __init__ reached, for a ~92% reduction).

Didn't see a meaningful performance impact on macOS (cached reads are cheap?). I'd expect a more noticeable positive change on Windows though.

This is additive to #3994, with respect to #3993.

Test Plan

No type-inference logic changes; tests return identical before/after test results.

[github-actions]

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

[yangdanny97]

I can try to find someone with a windows machine to do some benchmarking