How is the EVA-CLIP-B tokenizer different from CLIP tokenizer? I am not able to find the details in the paper
