Efficient Adversarial Training via Criticality-Aware Fine-Tuning