bland sum along axis (statistical_launch_wrapper) gives swapped result

[n-west]

When looking at the pfb shape generation I noticed doing it on the gpu results in a shape that seems shifted (like fftshift). This only seems to happen when the sum step is done with the cuda kernel. Each other step can be done on cpu or this step can be done on cpu at the result is as expected. Look in to the striding of the outputs matching input striding.

[n-west]

I think #71 was meant to capture a similar or the same issue and I may have identified the source is from laziness on my end inside some of the more complex cuda operators.

I noticed while working on #89 that the statistical operators weren't using the offsets to update pointers. I fixed at least part of the issue there and this may need to be looked at more carefully to get a systematic fix then test the gamut of potential usages