Evaluating point-light biological motion in multimodal large language models