Understanding and Mitigating Overfitting in Prompt Tuning for Vision-Language Models