OmniVL: One Foundation Model for Image-Language