End-to-end Reinforcement Learning for Time-Optimal Quadcopter Flight