CiT: Curation in Training for Effective Vision-Language Data