Altered arcTo's canonical points to (usually) be convex

BEFORE

arbroundrect  
8888: cmsecs =  48.52   
565: cmsecs =  46.33   
GPU: cmsecs = 125.11 gmsecs = 121.22  
NULLGPU: cmsecs =  58.97 gmsecs =   0.12

AFTER 

arbroundrect  
8888: cmsecs =  41.03   
565: cmsecs =  38.53   
GPU: cmsecs =  47.11 gmsecs =  45.97  
NULLGPU: cmsecs =  33.85 gmsecs =   0.11

[TomH, 2012-10-17 17:50:17]

Driveby LGTM

[bsalomon, 2012-10-17 19:26:26]

LGTM

https://codereview.appspot.com/6709051/diff/1/bench/PathBench.cpp
File bench/PathBench.cpp (right):

https://codereview.appspot.com/6709051/diff/1/bench/PathBench.cpp#newcode744
bench/PathBench.cpp:744: continue;
--i?

[reed1, 2012-10-17 19:38:22]

Sure, but...

More dox in the code
1. why its broken
2. why this works-around-that
3. other possible solutions for us to try
4. unittest(s)

[robertphillips, 2012-10-18 13:54:46]

PTAL

Note: this CL will alter the baselines for:
                degeneratesegments
                shadertext
                shadertext2

[reed1, 2012-10-18 15:16:43]

lgtm

[robertphillips, 2012-10-18 15:26:25]

committed as r5997

[robertphillips, 2012-10-19 12:43:53]

This breaks the following 21 Webkit tests

    clip-path-restore.html
    background-leakage-transforms.html
    background-leakage.html
    border-radius-split-background.html
    border-radius-circle.html
    border-shadow-large-radius.html
    border-styles-split.html
    border-radius-big.html
    box-shadow-clipped-slices.html
    shadow-tiling-artifact.html
    spread-multiple-inset.html
    spread-multiple-normal.html
    spread.html
    box-shadow-and-border-radius.html
    shadow-multiple.html
    shadow-multiple-vertical.html
    shadows.html
    border-styles-vertical-lr.html
    border-styles-vertical-rl.html
    border-radius-style-005.htm
    stroke-width-large.svg