Kernel normalization treatment #4

Untangle kernel re-normalization & NaN interpolation

• The following seems untrue, unnecessary, and inefficient (src)

 # Because the Cython routines have to normalize the kernel on the fly, we
 # explicitly normalize the kernel here, and then scale the image at the
 # end if normalization was not requested.

Hmmm, maybe not. The interpolation scheme seems to prevent this hoist and is the reason that the kernel summation exists in the Cython code. If the image element is NaN then the kernel element is not incorporated in the sum.

A silver lining is that the above is only true for nan_treatment = 'interpolate'. It can, therefore, be optimized out for all other options. NaN interpolation is, however, the default.

This can also be checked before convolving to determine whether NaN values even exist.