Fix: Refactor DelegatingMultiprocessQueueFactory for fork safety

[rwkeane]

This commit refactors DelegatingMultiprocessQueueFactory and its associated sink/source wrappers to address a FileNotFoundError that occurred when using the 'fork' multiprocessing start method with PyTorch tensors.

Key changes:

1. Eager Queue Creation:

   • DelegatingMultiprocessQueueFactory now eagerly creates two sets
     of underlying queues in its constructor (default and Torch-specific
     if PyTorch is available). This ensures queue resources are
     initialized in the parent process before forking.

2. Lazy Queue Selection:

   • DelegatingMultiprocessQueueSink and Source now use a
     coordination mechanism on the default queue. The first item's type
     determines if the Torch or default queue pair is used for actual
     data transfer for that and all subsequent items.

3. Unit Test Updates:

   • The test suite in delegating_multiprocess_queue_factory_unittest.py
     has been extensively updated.
   • The 'fork' + tensor test (test_mp_correctness_fork_tensor_item_first)
     now passes.
   • A new minimal test (test_minimal_torch_mp_queue_spawn_tensor)
     was added to help diagnose issues with 'spawn' and PyTorch tensors.

Outcome:

• The primary goal of fixing the FileNotFoundError with the 'fork'
  start method and PyTorch tensors is achieved. All 'fork' method
  tests, including tensor scenarios, now pass.
• 'Spawn' method tests with non-tensor items also pass.

Investigation of 'spawn' + PyTorch Tensor Failures:

• Extensive investigation was conducted into why tests using the 'spawn'
  start method with PyTorch Tensors fail with EOFError/FileNotFoundError
  (tests: test_mp_correctness_spawn_tensor_item_first,
  test_mp_correctness_spawn_mixed_items_tensor_first, and the
  diagnostic test_minimal_torch_mp_queue_spawn_tensor).
• The minimal test, which uses torch.multiprocessing.Queue directly
  (bypassing all tsercom factory/wrapper logic), also fails under
  'spawn' with tensors.
• Experiments included:
  • Explicitly setting PyTorch's sharing strategy to 'file_system'.
  • Explicitly calling tensor.contiguous().share_memory_() before
    putting tensors into the raw torch.multiprocessing.Queue in
    the minimal test.
• None of these experiments resolved the failures in the minimal test
  case for 'spawn' + tensor.

Conclusion on 'spawn' + PyTorch Tensor Issue:

• The persistent failures in the minimal torch.multiprocessing.Queue
  test strongly indicate that the root cause is external to the
  tsercom.DelegatingMultiprocessQueueFactory code. It is likely an
  underlying issue or limitation within PyTorch's multiprocessing
  module regarding tensor sharing with the 'spawn' start method in the
  current testing environment.
• The tsercom code now correctly handles 'fork' scenarios and 'spawn'
  for non-tensor data. The spawn+tensor issue is documented via the
  failing minimal test included in the test suite.