copy of the texels in VRAM rather than a readback and re-upload. This
gives a 3-10X speedup on recursive canvas-to-canvas draws.
N.B.: This introduces a new GM test, which will need new baselines.
git-svn-id: http://skia.googlecode.com/svn/trunk@2790 2bbb7eff-a529-9590-31e7-b0007b416f81
