Fix multiprocessing segfaults on Windows (#727)

[lanpa]

Description

This PR addresses the unpredictable segfaults on Windows reported in issue #727. The root cause was identified as unsafe pickling of SummaryWriter objects and background threads across processes, combined with child processes incorrectly closing shared resources during cleanup.

Changes:

• Serialization Improvement:
  • Event objects are now serialized to bytes before being put into the multiprocessing.Queue. This avoids complex pickling of Protobuf objects, which is often unstable across process boundaries on Windows.
  • Updated the background logger thread to handle both bytes and Event objects for backward compatibility.
• Multiprocessing Safety:
  • Added __getstate__ and __setstate__ to SummaryWriter, FileWriter, and EventFileWriter to explicitly exclude non-picklable resources (threads, file handles, and loggers) from being transmitted to child processes.
  • Resources are lazily re-initialized in child processes only if needed.
• Cleanup Refinement:
  • Added a PID check to the atexit cleanup handler to ensure that only the process that originally created the FileWriter can close it. This prevents child processes from prematurely closing shared queues and file handles.
  • Moved the cleanup handler to a top-level function to ensure it can be correctly pickled in spawn mode.

Verification Results:

• ✅ New Test Case: Added tests/test_issue_727.py which specifically tests SummaryWriter usage in a spawn-based multiprocessing environment (simulating Windows behavior on other platforms).
• ✅ Local Verification: Verified that the changes prevent the reported resource-closing crashes and that data is correctly recorded from worker processes.