TEST: Join SQLite cleanup workers before leak checks

[Richard Lundeen (richlundeen)]

Description

Fix the intermittent GC assertion in test_sqlite_instance_does_not_retain_process_exit_cleanup. It failed in the Windows Python 3.13 dev merge-group job 118365076379 while checking #3206.

dispose_engine_async awaits run_legacy_sync_async(self._dispose_sync_engine), which uses asyncio.to_thread. A thread-pool future can notify the event loop before its worker deletes the work item. That work item still holds the bound disposal method and thus the memory instance. gc.collect() cannot collect a live reference. This is a test lifetime race, not a process-exit hook leak or a result of #3206's seed-storage changes.

Run the existing fixture lifecycle in its own asyncio.Runner. Exit the runner to join its default executor before the unchanged GC and weak-reference assertion. Keep the checks that the real cleanup method is restored. No production code, shared fixture, timeout, retry, or leak assertion changes. This does not duplicate #3361 or #3344.

Tests and Documentation

• uv sync --frozen --extra all — passed; lockfile unchanged.
• uv run --no-sync pytest tests\unit\memory\test_sqlite_fixtures.py -q — passed, 14 tests on Windows Python 3.14.2.
• $env:UV_NO_SYNC='1'; uv run --no-sync pre-commit run --all-files — all hooks passed.
• git --no-pager diff --check — passed.
• Deterministic in-process race harness: pause ThreadPoolExecutor immediately after _WorkItem.run completes; release the worker when executor shutdown starts. The old GC assertion failed 10/10 times. The corrected test passed 10/10 times. Direct referrers included the bound disposal method and its coroutine.
• Leak-detection control: register a real bound disposal method with atexit after disposal. The corrected test still fails its weak-reference assertion, as required. The harness unregisters the callback afterward.

No documentation or notebook behavior changes; JupyText was not run. The full unit suite and the exact CI Python 3.13 environment were not run locally.