Support all_reduce a list of same-device tensors

[mrshenli]

distributed_c10d.py provides APIs to reduce a single tensor per process, or multiple tensors per process where each of them needs to reside on a different device. It will be useful (e.g., for model averaging) to support reducing a list of tensors per process where all of them are on the same device. The implementation should apply bucketing under the hood to get better performance.

[pietern]

This wouldn't allreduce them together (as is the case for the multi device one) but rather concatenate them and allreduce them in a single call, as opposed to N separate calls. We can call this allreduce_coalesced, similar to the existing broadcast_coalesced.

[pietern]

This has been added in #24949 thanks to @jfc4050!