Open-Vocabulary Skeleton Action Recognition with Diffusion Graph Convolutional Network and Pre-Trained Vision-Language Models.