VLPD: Context-Aware Pedestrian Detection via Vision-Language Semantic Self-Supervision