An Investigation into Pre-Training Object-Centric Representations for Reinforcement Learning