Improve idist for gather using nccl and reduce for gloo + gpu

🚀 Feature

Consider the following piece of code

def write_preds_to_file(predictions, filename):
    prediction_tensor = torch.tensor(predictions)
    prediction_tensor = idist.all_gather(prediction_tensor)

    if idist.get_rank() == 0:
        torch.save(prediction_tensor, filename)

The idist.all_gather() is used to collect the tensor from all the processes even if only the rank 0 needs it. The gather() method would be used but the backend nccl does not support it. See here.

The idea here is to implement the gather() method in idist using all_gather() for nccl (and gather() for others backends). Note that reduce() for gloo on GPU could be implemented using all_reduce() in a similar way.

It needs tests + docs

Start by locating the idist implementations of gather and reduce and review how they select distributed backends. Add the requested backend-specific behavior for NCCL and Gloo on GPU, then add tests and documentation as noted in the issue; done means both operations work for the described backend combinations.