Instead of checking just the monotonicity of the blending results, this
changes the alpha-blending test to compute the reference blend result
itself and then comparing to the compositor result. This way we can be
sure that the compositor implements the exact correct formula and not
something that just looks nice, as verifying the reference images are
actually correct is hard.

The reference image is renamed to follow the fact that this is not
primarily a monotonicity test anymore. The reference image is also
redundant, but I think it has documentary value.

The #if 0'd block of code was very useful in figuring out blending
errors in a future test case, so it is included here. I have a feeling
we are going to need it again.

Signed-off-by: Pekka Paalanen <pekka.paalanen@collabora.com>