Visually Grounded Self-Reflection for Vision-Language Models via Reinforcement Learning