2.17.26 regression tests

[Keith OHara]

I re-pushed the fix for issue 2910, which fixes the implementation of 
outside-staff-padding, so many many regtests change slightly ... for the third 
time.  Sorry about that.

(In hindsight, I should not have reverted the original 2910 fix.  The first 
report of a really bad MSWindows-only bug happened to come while verifying the 
fix for 2910.)

I looked through the automatic comparison at 
http://lilypond.org/test/v2.17.26-1/compare-v2.17.25-1/index.html, and 
recognize causes for all the changes, though it is hard to keep concentration 
through the long list of changes.

Maybe we 'dynamics-avoid-cross-staff-stem.ly' looks a bit worse, because the 
(erroneous) double value of outside-staff-horizontal-padding was keeping the p 
and fff apart.  Similarly, the text labels in 'accidental-ancient.ly' now slide 
next to each other because they just happen to fit perfectly over the music.

We could increase the default padding so that it has more effect without being 
(erroneously) doubled;  (I didn't think that increase was warranted overall or 
I would have done it.)  People did seem to like objects getting closer in 
similar situations, with the original stencil-integral patch.

The test 'parenthesize.ly' now fails to meet its self-description, but that was due a 
change to the test input by David, so it seems to be one of his 
"not-yet-regression" tests that will pass after the patch for issue 1523 goes 
in.

We have a release with no known regressions.