Spotlight: Object Abstraction in Visual Model-Based Reinforcement Learning