Tree of Attributes Prompt Learning for Vision-Language Models