Why Large Language Models and Humans Converge and Diverge in Evaluating Creativity