LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models