Blender Git Loki

Deep shadow buffers are now created and stored tiled, this helps avoid a
memory peak when creating them, and also can save some memory afterwards
by not storing empty tiles. Threading code here was also changed so that
each tile can be rendered by it's own thread, and memarena is now used
to avoid slowness due to malloc mutex locks. There's no setting at the
moment to configure the tile size, fixed to 256x256 currently to avoid
performance problems.