Gradient-Regulated Meta-Prompt Learning for Generalizable Vision-Language Models