Towards Unified Multi-Modal Personalization: Large Vision-Language Models for Generative Recommendation and Beyond