Motif-Consistent Counterfactuals with Adversarial Refinement for Graph-Level Anomaly Detection