Streamline picture_record_recurring_ bench.

Streamline picture_record_recurring_ bench.

Reduce benchmark runtime from ~17 ms to ~13.25ms by moving SkPaint
construction out of the inner loop and by using random colors in the
SkPaints instead of sequential integers, which were seeing bad hashing.

This doesn't actually improve our performance, but makes the benchmark
more focused on pure recording costs.

BUG=

Committed: http://code.google.com/p/skia/source/detail?r=10003

[tomhudson, 2013-07-11 10:19:50]

When we're running the picture_record_recurring_paint_dictionary benchmark, we
spend a lot of time in the SkPaint constructor, and we get a very high rate of
hash table conflicts (it seems to me that only 16 of 128 hash bins are used).

This change removes SkPaint construction from the inner loop and uses
non-sequential colors for the paints, reducing benchmark time about 20% on
Linux. The benchmark is now only timing canvas->drawPaint(). This isn't an
actual performance win, but it makes the benchmark more focused on the cost of
picture recording. Do we want that increased focus?

[reed1, 2013-07-11 13:49:20]

lgtm

very good change -- focus in one what we care about

[mtklein, 2013-07-11 13:56:21]

lgtm

[commit-bot: I haz the power, 2013-07-11 14:00:37]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/tomhudson@chromium.org/18119011/1

[commit-bot: I haz the power, 2013-07-11 14:26:11]

Change committed as 10003