Improve wcs comparison performed by compare_asdf in regression tests.

[braingram]

As mentioned here: #868 (review)

The scientifically relevant WCS consistency is about ~1/100 pix, which would correspond to 0.001 arcsec atol. One might hope that for this kind of regression testing we could do 10 or 100x better than that. That would be in 2D, so it would be something like np.all(coord1.separation(coord2) < 0.001 * u.arcsec).

As WCS objects don't implement an equality operator #868 projects the wcs bounding box using each wcs (one from the truth, one from the result file) and compares the output. Applying the above suggestion would make this wcs comparison more closely match the scientific needs of the data files.