ANAH: Analytical Annotation of Hallucinations in Large Language Models