Improve dithering in 32bit gradients

Go from a 1x2 to a 2x2 dither cell, and change/simplify the logic for how we
compute the "dithered" version of a color to just a graduated fixed-point-round.
Also, add this new dither to conical and sweep, which before had no dithering.

Disabled for now using SK_IGNORE_GRADIENT_DITHER_FIX. Will enable this and
and rebaseline skia.

http://code.google.com/p/skia/issues/detail?id=1098
Committed: https://code.google.com/p/skia/source/detail?r=7549

[reed1, 2013-01-30 22:44:03]

I think this CL is better than the existing code, but having now see what Safari
draws, I'm not sure if we should commit this now (and rebaseline), or first try
to improve quality overall to match Safari.

[junov, 2013-01-31 14:42:17]

Do you have any ideas or leads for improving quality?

[junov, 2013-01-31 14:45:59]

If you have to reduce the amplitude by half to make the problem go away, I think
there might be a systematic bias in the computation somewhere (Bad rounding or
something like that).

[caryclark1, 2013-01-31 14:55:41]

lgtm

but I agree with the CG quality sentiment.

[reed1, 2013-01-31 15:08:17]

I have some experiments in play to try to improve quality, though they all
involve some perf hit and possible some ram hit.

I won't try to land this now, given the massive rebaseline, until I spend more
time trying to improve the quality... to a point. Rebaselining is big, but not
actually hard, and if we have to do it more than once (as we iterate on quality
improvements), that's not the end of the world.

[reed1, 2013-02-01 21:39:56]

ptal

[reed1, 2013-02-01 21:50:32]

#5 fixes typo in hex-offsets, found by cary

[junov, 2013-02-01 21:54:03]

On 2013/02/01 21:39:56, reed1 wrote:
> ptal

Yes! lgtm
How does quality compare?

[reed1, 2013-02-01 21:59:03]

On 2013/02/01 21:54:03, junov wrote:
> On 2013/02/01 21:39:56, reed1 wrote:
> > ptal
> 
> Yes! lgtm
> How does quality compare?

http://code.google.com/p/skia/issues/detail?id=1098

The images attached to this bug show some screenshots.