Understanding Self-Predictive Learning for Reinforcement Learning