FR Transpose transformation

[akern40]

Issue Description

Many normalizing flows require the use of transpositions during the flow, allowing different transformations to operate on different dimensions. Since transpositions don't change the volume of a tensor, this operation would be relatively easy to implement.

One question of implementation would be how to handle transpositions on batches, i.e.

x = torch.rand(5, 4, 3)

# If we only want to permute the last two dimensions, can we do
transpose = Transpose((0, 1))
transpose(x).shape  # Would show [5, 3, 4]

# Or do we have to fully specify the entire thing
transpose = Transpose((0, 2, 1))
transpose(x).shape  # Would show [5, 3, 4]

While the first method seems a little odd, the advantage is that you can handle both batched and un-batched tensors (the transformation would assume the transposition happens on the rightmost dimensions).

cc @stefanwebb

[akern40]

After some quick testing, turns out negative indices work with torch.permute, so perhaps the best thing to do would simply be to require negative indices be passed to the transformation:

x = torch.rand(5, 4, 3)

# Using negative dimensions
transpose = Transpose((-1, -2))
transpose(x).shape  # Would show [5, 3, 4]

This still makes an assumption about operating on the rightmost tensor dimensions but could make the assumption more clear to the end-user. The expanded permutation could be achieved by prepending zeros to the given all-negative permutation.