PyTorch's new PR enables CUDA symmetric memory without NCCL

PyTorch merged PR #184523, which adds test coverage for CUDA symmetric memory (symm_mem) when backed by torchcomms through the new _BackendWrapper shim. The PR introduces a test class `TorchCommsCudaSymmMemTest` that validates two rendezvous strategies: one where metadata allgather flows through the torchcomms-backed ProcessGroup, and another where metadata exchange falls back to the default TCPStore. In both cases, the test allocates a symmetric memory buffer, rendezvouses on the ProcessGroup, and verifies each rank can read its peers' buffers.

This test serves as both a regression guard and an example of how to use the _BackendWrapper with CUDA symmetric memory. By enabling symm_mem without requiring ProcessGroupNCCL, PyTorch opens the door to more flexible distributed training setups, especially in heterogeneous environments or when using custom communication backends. The PR was approved by ngimel and fduwjj, key maintainers of PyTorch's distributed module.