Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models