DHCP: Detecting Hallucinations by Cross-modal Attention Pattern in Large Vision-Language Models