VLCDoC: Vision-Language Contrastive Pre-Training Model for Cross-Modal Document Classification