MindGYM: Enhancing Vision-Language Models via Synthetic Self-Challenging Questions