BEiT v2: Masked Image Modeling with Vector-Quantized Visual Tokenizers