This filter performs 'reflection' or 'mirrored' padding, e.g.
[1, 2, 3] with paddings [2, 1] (before, after) becomes
[3, 2, 1, 2, 3, 2]. The boundaries are not repeated, which
resembles TensorFlows behaviour.
The current implementation is large and rather slow, adding
~2 seconds overhead to a full style model run for the default
sample image on my machine.
x: <1,2,4>D tensor
paddings: number[[number, number]] - Before/after pad for each dim