Patch-Prompt Aligned Bayesian Prompt Tuning for Vision-Language Models