A Geometric Perspective on Optimal Representations for Reinforcement Learning