Flying on Point Clouds with Reinforcement Learning