In practice, the A and B images may be any mixture of RGB24 and
ARGB32 formats, so this change accepts all combinations of these
types, and converts the pixel values to a common (ARGB32) format
as needed.

Some of the newly added test failures are cases where the image
output is RGB24, but the matching reference image is ARGB32 with
noticeable transparency. Some of the newly passing tests are cases
where the unused 'alpha' channel of an RGB24 image was not equal
to 0xff, and the previous code had incorrectly used this channel
in max_diff calculations.