VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning