“Dmitri Dolgov believes a pure "pixels-in, trajectories-out" end-to-end model is insufficient for achieving the high level of safety and performance required for a fully autonomous system.”