Towards Provable And Practical Reinforcement Learning With Representation Learning