ClipGrader: Leveraging Vision-Language Models for Robust Label Quality Assessment in Object Detection