Fixes the "nonzero is not supported for tensors with more than INT_MAX elements" in get_all_triplets_indices

[mkmenta]

This solves #688

[mkmenta]

I am not sure if there could be other places where changing the torch.where could be beneficial (in case this implementation is accepted).

I think that in get_all_pairs_indices it would be hard to have a tensor with INT_MAX elements because the tensors used in the torch.where in this case are just len(labels)*len(ref_labels).

[KevinMusgrave]

Thanks!