Learning Accurate Long-term Dynamics for Model-based Reinforcement Learning