Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning