Machine learning potentials for metal-organic frameworks using an incremental learning approach