BUG: make log rotation safe across Windows processes

[Ricardo-M-L]

Summary

• use msvcrt.locking on Windows while retaining fcntl.flock on POSIX
• serialize Windows writes with rollover so no record can land between the copy and truncate fallback steps
• close the lock file if native lock acquisition fails
• persist a rotation generation and timed-rollover marker so waiting processes do not rotate or overwrite the same archive again
• share the coordination logic across size, timed, and combined handlers

Fixes #5284.

Validation

• xinference/deploy/test/test_log_rotation.py: 37 passed
• repository pre-commit hooks: Black, Ruff, isort, mypy, codespell, whitespace checks all passed
• rebased onto current upstream/main

The regression coverage includes a deterministic concurrent-write test that pauses after the Windows fallback copy, verifies a sibling writer blocks, and confirms the record is preserved after truncation. Windows locking behavior is simulated on macOS; the repository Windows CI job provides the native-platform check.

[gemini-code-assist]

Code Review

This pull request introduces cross-platform, cross-process log rotation safety by implementing a shared _SafeFileRotationMixin that uses platform-native file locking (msvcrt on Windows and fcntl on Unix) and tracks rotation state via a shared lock file. It updates the rotating file handlers to inherit from this mixin and adds comprehensive unit tests to verify the Windows fallback and rotation coordination. The feedback highlights a potential file descriptor leak in _acquire_rotation_lock if the locking mechanism raises an exception, suggesting wrapping the lock acquisition in a try...except block to ensure the file descriptor is closed.

[Ricardo-M-L]

Friendly ping - this PR makes log rotation safe across Windows processes. Would appreciate a review. Thanks!

[qinxuye]

Please also address the unresolved Gemini review comment: _acquire_rotation_lock() must close lock_fd if msvcrt.locking() or fcntl.flock() raises.

[qinxuye]

LGTM. Verified both previous findings are fixed: failed lock acquisition closes the descriptor, and Windows writes share the rotation lock across copy/truncate. Both threads are resolved. All 37 focused log-rotation tests passed locally, including the simulated Windows concurrent-write regression; CI is green.