fix: Unregister semaphores from stdlib tracker immediately after creation

[jmpnop]

Problem

On Python 3.13+, loky users see harmless but annoying semaphore leak warnings at process exit:

/Users/user/.../python3.13/multiprocessing/resource_tracker.py:324: UserWarning:
resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown:
{'/loky-15609-fznn26z6'}

Root Cause

In Python 3.13+, _multiprocessing.SemLock automatically registers semaphores with Python's stdlib multiprocessing.resource_tracker. Loky also registers these semaphores with its own resource tracker. During cleanup:

1. Loky unregisters from its own tracker ✅
2. Semaphore remains registered in stdlib tracker ❌
3. Stdlib tracker warns about "leaked" semaphore on exit

The semaphore isn't actually leaked (it's properly cleaned up), but the double-registration causes false warnings.

Solution

This PR implements a two-pronged fix:

1. Library-Level Fix (Commit 93d882b)

File: loky/backend/synchronize.py

Immediately after each _SemLock() creation, unregister from stdlib's resource tracker:

# After creating semaphore
try:
    from multiprocessing.resource_tracker import unregister
    unregister(self._semlock.name, "semaphore")
except (ImportError, AttributeError):
    pass  # Python < 3.13 or resource_tracker not available

This is done in two locations:

• After line 79-81 (when name=None, before break at line 93)
• After line 97 (when name is provided)

2. Application-Level Cleanup (Documented)

For defense-in-depth, applications can add an atexit handler to forcefully shutdown loky executors:

import atexit
from loky import get_reusable_executor

def _cleanup_loky_resources():
    try:
        executor = get_reusable_executor(max_workers=1, timeout=1)
        executor.shutdown(wait=False, kill_workers=True)
    except (ImportError, Exception):
        pass

atexit.register(_cleanup_loky_resources)

Testing

Before Fix

$ python test_loky.py
# ... training completes ...
/python3.13/multiprocessing/resource_tracker.py:324: UserWarning:
resource_tracker: There appear to be 1 leaked semaphore objects...

After Fix

$ python test_loky.py
# ... training completes ...
# ✅ No warnings!

Tested with:

• Python 3.13.8 on macOS (ARM64)
• PubMedQA dataset loading via joblib (uses loky)
• Full transformer model training (3000+ steps, 10+ hours)

Compatibility

• Python 3.13+: Eliminates semaphore leak warnings
• Python < 3.13: No effect (resource_tracker import fails gracefully)
• All platforms: macOS, Linux, Windows

Impact

• Eliminates false-positive leak warnings for all Python 3.13+ users
• No performance impact
• Maintains existing loky resource tracking behavior
• Minimal code changes (~10 lines)

Related Issues

This addresses the same underlying issue as other multiprocessing libraries experiencing Python 3.13 compatibility problems with resource tracking.