Efficient Model-free Reinforcement Learning in Metric Spaces