Humans and LLMs Diverge on Probabilistic Inferences