AD-LLM: Benchmarking Large Language Models for Anomaly Detection