Support `dist.ReduceOp.AVG` on XLA device

[miladm]

🚀 Feature

Enable dist.ReduceOp.AVG on XLA. This operation is available on NCCL devices only. (Reference). This op enables operations like torch.distributed.all_reduce(<tensor_name>, op=dist.ReduceOp.AVG)

Error when using the op

ValueError: Invalid reduce op <torch.distributed.distributed_c10d.ReduceOp object at 0x7f0cb2d3b830>