[cairo] New regressions since 10/17

[Daniel Amelang]

> I checked the images and they new failures differ by a maximum change
> of 1 in any color channel...I neglected to install the new
> reference images with the change.

Along these lines...are you aware of PerceptualDiff? It uses "a
computational model of the human visual system to compare two images"
with the purpose of reducing false negatives in the case where the
images are not pixel identical, but the difference is imperceptible to
a human. You can tweak all sorts of parameters to make it more or less
'picky'. Supposedly Pixar uses it to track regressions in some
rendering software of theirs (the author worked for Pixar). More here:

http://pdiff.sourceforge.net/

I was playing around with it the other day, and got halfway into
incorporating it into the image_diff function currently found in
cairo-test. If you think it would help, I'll finish it up and submit a
patch.

Dan