[ROCm] rccl performance improvement via env var (#76985)
The env var HSA_FORCE_FINE_GRAIN_PCIE=1 enables P2P communication in RCCL without intermediate buffers. This is necessary on hosts with only PCIe and no P2P high-speed interconnect.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76985
Approved by: https://github.com/ezyang
