ProcessGroupNCCL error/timeout handling

[pietern]

As of NCCL 2.4 there are functions to detect I/O errors and abort running kernels. This is required to implement timeouts and force workers to raise an error or terminate when other workers fail.

See https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/api/comms.html and ncclCommAbort and ncclCommGetAsyncError and https://devblogs.nvidia.com/massively-scale-deep-learning-training-nccl-2-4/ for an example.

[pritamdamania87]

@pietern Should we close this out? @mrshenli created a follow up for error handling within the server: #24906.

[pietern]

@pritamdamania87 Yep, sounds good.