Add special handling of rectori case for gpu

I'm a bit disappointed in the performance improvement but the blurs still appear to dominate (in both the rectori GM and theverge skp).

GMs (old on left; new on right; everything <= .5ms || <= 3% || not appearing in two passes with repeat=100 removed):

-----
     ovals_3_stroke_4 GPU c          9.20      10.14      -0.94   -10.2% 
           tablebench GPU c         23.24      24.02      -0.78    -3.4% 

              rectori GPU c         81.59      70.36     +11.23   +13.8% 
-----

For bench_pictures (with same removals as with the GMs):

-----
     desk_yahoonews.skp tile_256x256_gpu c     69.26      66.14      +3.12    +4.5% 
           tabl_mlb.skp tile_256x256_gpu c     49.30      46.64      +2.66    +5.4% 
        tabl_nofolo.skp tile_256x256_gpu c     13.26      12.48      +0.78    +5.9% 
      tabl_theverge.skp tile_256x256_gpu c    211.38     193.44     +17.94    +8.5% 
          tabl_digg.skp tile_256x256_gpu c     49.92      43.06      +6.86   +13.7% 
-----

[reed1, 2013-05-16 14:13:20]

1. style nit on SkPath.h
2. unittest for the new call on path?

https://codereview.chromium.org/15080010/diff/2001/include/core/SkPath.h
File include/core/SkPath.h (right):

https://codereview.chromium.org/15080010/diff/2001/include/core/SkPath.h#newc...
include/core/SkPath.h:590: bool isNestedRects(SkRect rect[2], Direction* dirs =
NULL) const;
Direction dirs[2] = NULL

[caryclark, 2013-05-16 14:38:02]

lgtm (w/unit test as Mike says)

[bsalomon, 2013-05-16 14:54:08]

ditto

[robertphillips, 2013-05-16 15:33:55]

PTAL - piggy-backed direction-returning unit test on top of existing unit test.

https://codereview.chromium.org/15080010/diff/2001/include/core/SkPath.h
File include/core/SkPath.h (right):

https://codereview.chromium.org/15080010/diff/2001/include/core/SkPath.h#newc...
include/core/SkPath.h:590: bool isNestedRects(SkRect rect[2], Direction* dirs =
NULL) const;
On 2013/05/16 14:13:20, reed1 wrote:
> Direction dirs[2] = NULL

Done.

[reed1, 2013-05-16 16:24:22]

lgtm

[Tom Hudson, 2013-05-16 17:25:07]

Is there any reason we couldn't write something like this for CPU?

[robertphillips, 2013-05-16 17:45:29]

Mike can correct me but I believe the raster path performs a 9-patch
optimization in this case. Additionally rendering the nested rects isn't an
issue for the raster path since it just sorts the edges and goes. For the gpu it
causes particular problems since (for concave AA paths) we call back to raster
to render it for us.

[robertphillips, 2013-05-17 12:50:55]

committed as r9175