ViT : An Image Is Worth 16x16 Words: Transformers For Image Recognition at Scale