Control-Oriented Model-Based Reinforcement Learning with Implicit Differentiation