SEAS: Self-Evolving Adversarial Safety Optimization for Large Language Models