A Single Transformer for Scalable Vision-Language Modeling