Add numpy-based bilinear downsampling along X and Y

[j6k4m8]

!!!!! THIS IS COMPLETELY UNTESTED because I have no way to run this in the boss infra. The added code all works as it says on the label, but I don't know how it plays in its Boss context.

This reimplements the XY downsample as a numpy kernel convolution implementation of the bilinear downsample we're currently using, rather than converting to and then back from an image.

Additional Warnings, in case the first wasn't scary enough:

• This ONLY downsamples down by a factor of 2. I think that's correct (we use data from res (n) to get to res (n+1), right?), but will BREAK if the volume and cube arguments are not the same z-size and if cube is not half of volume in X and Y.
• I hand-wrote this kernel, so I think it's right but I might be mistook

Good things:

• This works on any datatype, so we support u8/u16, and also any other datatype
• Because we're not converting to/from an Image buffer, I think this actually saves us some time (?)

[sandyhider]

I am not going to merge it until I test on my stack.