Improve the curve tessellation subdivision of the path renderers.

Improve the curve tessellation subdivision of the path renderers.

The stencil-and-cover and tessellating path renderers currently
tessellate curves to linear segments to within one screen space pixel.
This is fairly poor accuracy, as can be seen from the strokecircle GM:
there are noticeable rotating bumps on the circles.

Improving the tolerance to 0.25 pixel eliminates the bumps, and
approximates the 16x supersampling of the raster path. 
This does have a performance hit: 3-6% on desktop on 
the IE chalkboard demo, ~1% on Nexus 7 2013.

Note: this will require rebaselining a number of GPU and  MSAA images 
in Skia, but nothing in Chrome (yet).

BUG=skia:3731
Committed: https://skia.googlesource.com/skia/+/2b4bb07492790d7d18dc046f05999915d4384c5d

[Stephen White, 2015-04-22 20:22:48]

senorblanco@chromium.org changed reviewers:
+ bsalomon@google.com

[Stephen White, 2015-04-22 20:23:49]

Brian: PTAL. Thanks!

[bsalomon, 2015-04-22 20:25:34]

lgtm

[Stephen White, 2015-04-22 20:29:47]

The CQ bit was checked by senorblanco@chromium.org

[commit-bot: I haz the power, 2015-04-22 20:30:05]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1072273007/20001

[commit-bot: I haz the power, 2015-04-22 20:45:22]

Committed patchset #2 (id:20001) as
https://skia.googlesource.com/skia/+/2b4bb07492790d7d18dc046f05999915d4384c5d

[tomhudson, 2015-04-23 17:16:23]

I think we're seeing 50-100% regression on Android for the tiger & map SVG test
cases. Is this really warranted?

https://perf.skia.org/#3251

[Stephen White, 2015-04-23 17:30:45]

On 2015/04/23 17:16:23, tomhudson wrote:
> I think we're seeing 50-100% regression on Android for the tiger & map SVG
test
> cases. Is this really warranted?
> 
> https://perf.skia.org/#3251

Sadly, yes. The quality is simply not acceptable with out (see the SVG tiger's
eye, for example). 

OTOH, there's a flaw in the way that the skia perf tests measure timings on the
GPU which overweights the cost of CPU work, and underweights the cost of GPU
work (I've been meaning to file a bug about this once I have the data.) So this
regression is not as bad as it looks, in terms of whole-system performance.