MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data