Blind Image Quality Assessment via Vision-Language Correspondence: A Multitask Learning Perspective