Optimizing Automatic Differentiation with Deep Reinforcement Learning