gh-150818: Wire logger parent before publishing it in getLogger() (GH-150941)

[gaborbernat]

The lock-free fast path added in GH-150825 reads loggerDict without the lock and returns any non-placeholder logger it finds. The slow path, however, inserted a freshly created logger into loggerDict before calling _fixupParents(), which assigns its parent. A caller hitting the fast path during that window receives a logger whose parent is still None: getEffectiveLevel() returns NOTSET and records do not propagate to ancestors until the creating thread finishes.

Both slow-path branches now publish the logger to loggerDict only after _fixupParents() (and _fixupChildren() for the placeholder case) have wired it. Neither helper reads loggerDict[name], so the move is safe, and the fast path is unchanged.

The added regression test forces the publish/wire interleaving deterministically and fails without this change.

The reorder is confined to the locked creation path, so the fast path is unaffected. Median of 5 runs against GH-150825 as merged (PGO build, same interpreter, swapping Lib/logging/__init__.py):

Benchmark	base (GH-150825)	patched
getLogger, 23 existing loggers	2.68 us	2.65 us: 1.01x faster

The difference is within the run-to-run noise floor (base 2.66-2.76 us, patched 2.64-2.65 us); pyperf compare_to reports it as not significant.

Benchmark (pyperf)

import logging, pyperf

names = ["elastic_transport.transport", "peewee.pool", "LiteLLM", "httpx",
    "uvicorn.error", "uvicorn.access", "fsspec", "boto3", "posthog", "amqp",
    "nox", "tldextract.cache", "dulwich.lfs", "urllib3.connectionpool",
    "asyncio", "sqlalchemy.engine", "werkzeug", "django.request", "celery.worker",
    "kafka.conn", "redis.client", "paramiko.transport", "matplotlib.font_manager"]
for n in names:
    logging.getLogger(n)

runner = pyperf.Runner()
runner.bench_func("getLogger %d existing loggers" % len(names),
                  lambda: [logging.getLogger(n) for n in names])

Base is Lib/logging/__init__.py at GH-150825; patched swaps in this PR's version on the same interpreter. Run base vs patched with pyperf compare_to base.json patched.json.

Thanks to @methane for catching this in review of GH-150825.

• Issue: Speed up logging.getLogger() for already-registered loggers #150818

[vsajip]

The new test is failing under Emscripten/WASI due to an inability to start a new thread - I'm not sure if this a transient problem or a platform restriction of those environments. It may be an idea to skip that test in those environments, at least for now.

[gaborbernat]

@vsajip not sure if that backport landed but if it did we either need to backport this too, or back out the backports 🤔

[vsajip]

No, as per discussion on d.p.o, the backport PRs were closed. Thanks for this! I took my eye off the ball a bit with the original PR.