A Unifying Framework for Action-Conditional Self-Predictive Reinforcement Learning