DVM: Towards Controllable LLM Agents in Social Deduction Games