Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in Multi-Agent RL