Seeing is Believing: Mitigating Hallucination in Large Vision-Language Models via CLIP-Guided Decoding