Exploiting Symmetry in Dynamics for Model-Based Reinforcement Learning with Asymmetric Rewards