The Value-Improvement Path: Towards Better Representations for Reinforcement Learning