recwarn: pass originating module name when re-emitting unmatched warnings (fixes #11933)

[mokashang]

Problem

pytest.warns re-emits unmatched warnings on __exit__ via
warnings.warn_explicit(..., module=w.__module__, ...). But w is a
warnings.WarningMessage instance, so w.__module__ evaluates to
"warnings" (the module that the WarningMessage class is defined in) for
every re-emitted warning, regardless of where the warning actually came from.

The visible consequence: a user-installed filter such as

warnings.filterwarnings("ignore", category=UserWarning, module=r"my_module")

matches the original warning fine, but silently fails to match its re-emit.
Code that should be quiet around pytest.warns(...) blocks gets noisy
(or, with "error" filters, raises unexpectedly).

This is #11933, reported by @Zac-HD with a precise diagnosis: module=w.__module__
should be the importable module name of the call site, not the home module of
WarningMessage.

Approach

WarningMessage does not preserve the module argument originally passed to
warn_explicit, so we can't read it off w. The next best thing is to look up
sys.modules by __file__ to find the module whose source file matches
w.filename, and fall back to None when nothing matches. warn_explicit's
documented behavior with module=None is to derive a name from filename
itself, which matches what plain warnings.warn(...) would have produced if
called from outside any pytest.warns context — so the fallback is a clean
degenerate, not a regression.

The fix is contained to two spots in src/_pytest/recwarn.py: a helper
_module_name_for_filename, and replacing the module=w.__module__ argument
in the re-emit call.

Why this PR re-opens ground covered by #12898

The earlier attempt (#12898, by @reaganjlee, closed for inactivity) landed
on essentially the same shape of fix but was closed because the regression test
@nicoddemus asked for was inadequate — specifically, the test passed even when
module= was set to None instead of the recovered name, so it pinned the
wrong invariant.

This PR addresses that directly. The new test
test_re_emit_preserves_module_name uses a packaged module
(regr_pkg/inner.py) so the importable name (regr_pkg.inner) differs from
warn_explicit's filename-derived default (<...>/regr_pkg/inner). The
filter regex ^regr_pkg\.inner$ matches only the importable form and
nothing else. Concretely, the test:

• passes with the fix (module recovered to "regr_pkg.inner"),
• fails with the original bug (module "warnings" doesn't match),
• fails with module=None (filename-derived "<...>/regr_pkg/inner" doesn't match).

So the test pins down "recover the correct module name", not the looser
"don't pass \"warnings\"".

Verification

• testing/test_recwarn.py — 67 passed (66 existing + 1 new)
• testing/test_recwarn.py testing/test_warnings.py — 119 passed, 4 skipped
• Full testing/ suite (excluding test_recwarn.py, kept separate above) —
  3974 passed, 124 skipped, 12 xfailed, 1 xpassed; no failures introduced.
• ruff check, ruff format --check, mypy on the touched files — all clean.

I also confirmed end-to-end by reverting the fix locally and re-running the
new test: it fails as designed. With the module=None "looser" fix it also
fails, which is what nicoddemus wanted from a regression test.

Fixes #11933.