SINC: Self-Supervised In-Context Learning for Vision-Language Tasks