Prompt Tuning Strikes Back: Customizing Foundation Models with Low-Rank Prompt Adaptation