Reward-Free Exploration for Reinforcement Learning