Fixes for mosaic output pixels not covered by inputs

I noticed that, when using reproject_and_coadd with background matching and with combine_function='first', a bounding box around some of the input images can be filled, so that outputs look like this: instead of this:

This happens because output pixels not covered by a given input image are set to zero after that image is reprojected, and those zero values get adjusted by the background matching. For the other combine_function options, those out-of-bounds pixels never get used, so their final value doesn't matter, but that wasn't the case for combine_function='first', which this PR rectifies.

While adding a test that works through the options and ensures the final output mosaics are zero outside the returned footprint and filled-in inside the footprint, I noticed an inconsistency in that all the combine_functions produce zeros for pixels not spanned by any input image, except for combine_function='mean', which outputs nans for those unsampled pixels (generated when dividing by sum-of-weight values of zero). This PR also changes those nans to 0, to ensure mean is consistent with the other modes.

astropy / reproject

Fixes for mosaic output pixels not covered by inputs #413

Codecov Report