Enhancing Adversarial Training via Reweighting Optimization Trajectory