Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?