Generalization and Robustness Implications in Object-Centric Learning