Disentangling What and Where for 3D Object-Centric Representations Through Active Inference