[Feature Request] Allow specifying nccl{Send,Recv} peer via CUDA tensor

[garrett361]

The current nccl{Send,Recv} API requires specifying a int peer for P2P ops. This causes unavoidable and undesirable device-host syncs in cases where the P2P op routings are dynamically determined.

Example: mixture of expert models (like DeepSeek v3) have dynamically-determined computation. The computation may require communication, depending on the parallelization scheme. The communication patterns can be encoded in, say, int64 CUDA tensors which contain the rank of the peers that different tensors should be sent do. Using the CUDA-tensor peer info to launch the corresponding P2P ops currently requires moving the CUDA tensors back to the host, so that the appropriate int peer values can be read off. This is an undesirable device-host sync point.

Request: allow specifying the peer via a pointer to a buffer held on the device.