Pre-training A Prompt Pool for Vision-Language Model