Growing a Multi-head Twig via Distillation and Reinforcement Learning to Accelerate Large Vision-Language Models