Characterizing the Role of Similarity in the Property Inferences of Language Models