Multi-Cache Enhanced Prototype Learning for Test-Time Generalization of Vision-Language Models