Segment Any Anomaly without Training via Hybrid Prompt Regularization