refactor torch/cuda/nccl.h to remove direct dependency on NCCL in libtorch_python (#42687)
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/42687
Reviewed By: malfet
Differential Revision: D23145834
Pulled By: walterddr
fbshipit-source-id: c703a953a54a638852f6e5a1479ca95ae6a10529