Instead of splitting the scene into multiple phases, use texture
layers to allocate slices as needed for each pass of the render
frame. These are pooled and only allocated when the frame render
target configuration changes.

Previously, the code asserted if it was unable to fit the render
tasks into a single render target for a single tile. Now, this
(pathological) case is handled correctly.

This makes the upcoming dynamic primitive cache support a lot
simpler to add. It also means that we don't use a ping-pong
allocation style, which typically performs badly on mobile / tiled
rendering architectures.