Is it good to add a sorted_input to torch.unique?

Current cuda implementation always sort the input tensor first. But sometimes the input tensor is guaranteed to be already sorted. Sorting is O(n*logn), which is likely to be the most time consuming part of the whole computation, and it would be nice if we allow user to skip this part.

torch.unique with sorted_input=True would resemble thrust::unique and only remove the consecutive duplicate elements.

It might be better to name it consecutive instead of sorted_input.