Use GrMemoryPool with GrCustomStage, add simple reclamation

Solves http://code.google.com/p/skia/issues/detail?id=736 by overriding new & delete for GrCustomStage and improving the performance of the memory pool for the access patterns we're seeing.

[TomH, 2012-07-24 19:23:36]

Without reclamation, we needed a pool size >> 100 objects to avoid thrashing.
With reclamation, we can get a pool size of 30 and never overflow into a second
block during the drawPosText benchmarks.

Hysteresis would also have helped, but reclamation should be more efficient when
it works.

[TomH, 2012-07-24 19:24:16]

Extra data point: without this, memory management made createTextureEffect ~ 15%
slower on the drawPosText benchmarks.

Using GrMemoryPool, we're paying maybe 1%.

[bsalomon, 2012-07-24 20:34:42]

http://codereview.appspot.com/6438046/diff/1/src/gpu/GrCustomStage.cpp
File src/gpu/GrCustomStage.cpp (right):

http://codereview.appspot.com/6438046/diff/1/src/gpu/GrCustomStage.cpp#newcode14
src/gpu/GrCustomStage.cpp:14: static GrMemoryPool gStagePool (30 *
sizeof(GrCustomStage), 30 * sizeof(GrCustomStage));;
Let's pick a byte count rather than base it on the size of the base class since
the base class is pretty minimal and most of the size will come from the
subclass.

http://codereview.appspot.com/6438046/diff/1/src/gpu/GrCustomStage.cpp#newcode52
src/gpu/GrCustomStage.cpp:52: //SkDebugf("custom stage new\n");
leave in?

[TomH, 2012-07-24 20:38:32]

After discussion:
1. 4k is reasonable
2. also need to remove static initializer and add thread-local storage

[TomH, 2012-07-24 21:36:30]

Committed in r4744 after verbal LGTM.