Energy-based Predictive Representations for Partially Observed Reinforcement Learning