In #22036 we added sparse allreduce for ProcessGroupGloo. It works for sparse CUDA tensors, but doesn't leverage InfiniBand like NCCL does. Therefore, we should have a sparse allreduce implementation for ProcessGroupNCCL as well.