Are language models rational? The case of coherence norms and belief revision