From chocolate bunny to chocolate crocodile: Do Language Models Understand Noun Compounds?