Mechanistic Anomaly Detection for "Quirky" Language Models