Towards Control-Centric Representations in Reinforcement Learning from Images