3. Fix leaked semaphore warnings on macOS (Python 3.10+)

[musicalplatypus]

Summary

Fixes the UserWarning: resource_tracker: There appear to be N leaked semaphore objects warnings that occur on macOS when training completes. This is caused by PyTorch DataLoader workers not being explicitly shut down before process exit.

Root Cause

On macOS with Python ≤3.11, multiprocessing.resource_tracker uses set.remove() instead of set.discard(), causing KeyError tracebacks when loky (scikit-learn's joblib backend) calls unregister() for semaphores registered in child processes.

Changes

1. Explicit DataLoader worker shutdown — Added ._shutdown_workers() calls in all train.py scripts (classification, regression, forecasting, anomaly detection) and all test_onnx.py scripts
2. Centralized cleanup in train_base.py — Created cleanup_data_loaders() utility and integrated it into the shared training base
3. Resource tracker unregistration — Added _unregister_semaphores() to properly clean up POSIX semaphores before process exit

Files Changed (13 files)

• tinyml-tinyverse/tinyml_tinyverse/references/common/train_base.py — core fix
• tinyml-tinyverse/tinyml_tinyverse/references/common/__init__.py — export
• 6× train.py scripts — per-task-type cleanup
• 5× test_onnx.py scripts — evaluation cleanup

Testing

• Verified on macOS 14 with Python 3.10, 3.11 — warnings eliminated
• No impact on Linux/Windows (fix is macOS-specific path)