[c10d] add ncclGetLastError to NCCL pg (#83724)
This PR add ncclGetLastError API to the nccl pg, to provide better error
reporting out of nccl failures directly, instead of guessing on random
reasons
Differential Revision: [D39161199](https://our.internmc.facebook.com/intern/diff/D39161199)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83724
Approved by: https://github.com/kwen2501, https://github.com/H-Huang