Provable and Practical: Efficient Exploration in Reinforcement Learning via Langevin Monte Carlo