Customizing Visual-Language Foundation Models for Multi-modal Anomaly Detection and Reasoning