Adds coverage for CUDASymmetricMemory rendezvous on a ProcessGroup whose backend is provided by torchcomms via _BackendWrapper, validating the shim path that lets symm_mem work without ProcessGroupNCCL.
Two variants are parametrized:
use_pg_for_symm_mem_rendezvous=True — metadata allgather flows through the torchcomms-backed...
🛡️ VERIFIED CYBER INTELLIGENCE ID: #3526650