Combating Adversarial Attacks with Multi-Agent Debate