The SPS allowed coding blocks down to 8x8. On NVIDIA that costs 4x the
encode time: 13.4 ms per 1920x1472 frame against 4.4 ms for the same
GPU's NVENC path, so the encoder sustained 36.5 fps against a 90 Hz
target. At 16x16 it is 3.2 ms and holds 90 Hz.
Replace the patches to Monado's main compositor with a local
implementation, reusing code when relevant.
The layer squasher is very close to comp_render_cs, while distortion is
completely replaced by foveation.