Objects matter: object-centric world models improve reinforcement learning in visually complex environments