VisPlay: Self-Evolving Vision-Language Models from Images