MetaVL: Transferring In-Context Learning Ability From Language Models to Vision-Language Models