Dynamics-Aware Comparison of Learned Reward Functions