Modeling Complex Behaviors: Multi-Personality Composition and Dynamic Switching in Vision-Language Models