Efficient Adversarial Training in LLMs with Continuous Attacks