An Information-Theoretic Approach to Reward Rate Optimization in the Tradeoff Between Controlled and Automatic Processing in Neural Network Architectures