Segmentation Downsampling

Hi Jeremy!

While working on Seung Lab's pipeline processing pipeline, I found that it's possible to use numpy to 2x2x1 downsample segmentation correctly for at least one mip level at reasonable performance. Would you be interested in a contribution? Two benefits are 1) Recursively downsampled segmentation mip levels won't visually march diagonally as they load. 2) MIP 1 is 100% accurate as in the most frequent pixel in a 2x2 grid is always chosen, so e.g. agglomeration algorithms can depend on it.

I have a 2x2x2 implementation as well that works in theory but that I haven't tested extensively.

I've written it up here: https://medium.com/towards-data-science/countless-high-performance-2x-downsampling-of-labeled-images-using-python-and-numpy-e70ad3275589

Thanks for the many enjoyable hours of neuroglancing. ^_^.

Will