This is implemented by using a GPU-accelerated surface for <canvas> when
a GPU context is available.
A side effect of this change is that all canvas modifications have to be
performed through Gfx::Painter, and whenever its content has to be
accessed, we need to take a snapshot of the corresponding GPU surface.
A new DrawPaintingSurface display list command is introduced to allow
cheap blitting of canvas content without having to read GPU surface
content into RAM.
