PyTorch adds multi-stream warning to prevent symm_mem deadlocks

PyTorch has merged a fix for a subtle but critical bug in its symmetric memory feature, which enables efficient multi-GPU communication. Collective kernels in `symm_mem` synchronize via a shared signal pad indexed by `blockIdx.x`. If two kernels from different streams are launched concurrently on the same collective group, they can deadlock silently—no error, no crash, just a hang. The root cause is that the shared signal pad cannot handle interleaved accesses from multiple streams without explicit synchronization.

To solve this, the PR (submitted by RohitRathore1 and approved by ngimel) adds a `warn_if_multi_stream()` helper that detects when collective ops are being used with multiple streams on the same group. It's wired into the affected operations, and the constraint is clearly documented in `symmetric_memory.md`. This prevents developers from accidentally violating the single-stream-per-group contract, turning a notoriously hard-to-debug silent hang into a clear, actionable warning. For teams running distributed training or inference on multi-GPU setups, this is a meaningful reliability improvement that reduces downtime and preserves the performance benefits of symmetric memory.