Object-Centric Scene Representations using Active Inference