PLPP: Prompt Learning with Perplexity Is Self-Distillation for Vision-Language Models