Edge treatment in convolution when edges are NaN is not consistent

[keflavich]

from astropy.convolution import convolve, convolve_fft, Gaussian2DKernel
import numpy as np
x = np.ones([100,100], dtype='float') * np.nan
convolve_fft(x, Gaussian2DKernel(2), fill_value=np.nan, boundary='fill')
convolve(x, Gaussian2DKernel(2), fill_value=np.nan, boundary='fill')

The behavior between the above functions is not consistent. convolve returns an image that still includes NaNs even though interpolate_nan is specified, while convolve_fft does not.

The convolve_fft version is naively more intuitive to me, but @adrn showed me a use case where the convolve behavior is (probably) what is desired.

We need to make these consistent and document how to achieve the different outcomes.