Rename "dummy" backend to "fake" and add init_process_group test

[d4l3k]

Summary

Rename the "dummy" backend to "fake" (matching PyTorch's FakeProcessGroup convention) and add a test that creates a process group using torch.distributed.init_process_group with the TORCH_DISTRIBUTED_USE_TORCHCOMMS=1 env var and the fake backend.

Good intro task — touches the build system, backend registration, and the PyTorch integration path.

Part 1: Rename "dummy" → "fake"

Rename TorchCommDummy → TorchCommFake and update all references:

• comms/torchcomms/TorchCommDummy.{hpp,cpp} — class, window, static registration name
• comms/torchcomms/tests/unit/cpp/DummyTorchCommBackend.cpp + CMakeLists.txt + TorchCommFactoryTest.cpp — dynamic loading test library and references
• comms/torchcomms/tests/unit/py/test_hooks.py, test_copy.py — new_comm("dummy", ...) calls
• setup.py — entry_points registration

Part 2: Add init_process_group test

Add a test that sets TORCH_DISTRIBUTED_USE_TORCHCOMMS=1 and calls torch.distributed.init_process_group with the fake backend. This exercises the upstream PyTorch integration path (distributed_c10d.py) where PyTorch itself creates the torchcomms communicator via new_comm(). Run a basic collective and tear down.

Build & test

USE_NCCLX=OFF USE_NCCL=OFF USE_TRANSPORT=OFF pip install --no-build-isolation -v .
ctest --test-dir build --output-on-failure   # C++
pytest comms/torchcomms/tests/unit/py/        # Python
lintrunner -a                                 # Lint