Distributional Reinforcement Learning for Energy-Based Sequential Models