Investigating Negation in Pre-trained Vision-and-language Models