Cache Session.gethookproxy during collection

[adamtheturtle]

Summary

• Cache Session.gethookproxy by path to avoid repeated conftest-set lookups during collection (addresses _pytest.main.Session.gethookproxy is called many times with the same argument on large suites pytest-dev/pytest#8991).
• Invalidate the cache when a new conftest module is registered.
• Do not cache a path until _loadconftestmodules has recorded its directory, so a premature FSHookProxy cannot stick after the directory is associated with already-imported conftests.

Typing of gethookproxy / ihook as HookRelay | FSHookProxy is intentionally not in this PR; see pytest-dev/pytest#14728.

Background (pytest-dev#2016)

gethookproxy(path) returns either the full hook relay or an FSHookProxy that strips conftests that do not apply to that path. An earlier path→proxy cache was removed in #2073 because it caused pytest-dev#2016:

Given siblings demo-0, demo-a (with pytest_ignore_collect → True), and demo-b:

1. Collecting root cached a “complete” proxy before any local conftest existed.
2. Entering demo-a loaded its conftest onto that same cached complete proxy.
3. Later collection under root reused the poisoned proxy, so demo-a’s ignore hook incorrectly affected siblings like demo-b.

The 2016 fix was to drop caching entirely and recompute every call. This PR restores caching for large-suite performance (pytest-dev#8991), but only with invalidation that covers that failure mode:

Guard	What it covers
Clear when len(_conftest_plugins) grows	Classic pytest-dev#2016: a newly imported conftest must not remain invisible (or wrongly global) on a cached proxy
Cache only if the directory is in _dirpath2confmods	Sibling case: path looked up before conftests are associated for that directory; association later does not grow _conftest_plugins, so generation alone would not clear a stale FSHookProxy

Timing

Warm interleaved A/B pytest --collect-only -q (alternate order). +% = faster.

Home Assistant (tests/, 122 476 tests, 3 pairs)

	Median
main	41.20s
This change	35.41s
Delta	+5.79s (+14.1%)

vws-python-mock (1 403 tests, 10 pairs)

	Median
main	0.315s
This change	0.375s
Delta	−0.060s (−19.0%)

On HA, gethookproxy runs ~233k times across ~8k unique paths, so cache hits dominate. On vws-python-mock (~2k calls / ~130 paths, ~0.2s collect) absolute savings are tiny and the measured delta is within run-to-run noise.

pytest’s own testing/ suite (~4k tests, 12 pairs)

Within noise (main 0.535s → 0.540s, −0.9%).

Test plan

• testing/test_pluginmanager.py::TestPytestPluginInteractions::test_hook_proxy (pytest_ignore_collect applies to all directories collected afterwards? pytest-dev/pytest#2016-style refresh after new conftest)
• testing/test_pluginmanager.py::...::test_hook_proxy_not_cached_before_directory_conftests_loaded
• testing/test_conftest.py::test_hook_proxy (sibling ignore_collect must not leak)
• testing/test_conftest.py::test_root_collect_file_applies_inside_package