ItemStudy: Added the use of large language models for judgment

Due to previous issues with semantic judgment, and the inability to simply use ReverseItem for judgment, we now call a large language model to determine whether the semantics of an item pair are opposite, and negate the similarity of item pairs with opposite semantics. Under a specific model (this time deepseek-chat), we selected both ends (similarity $>0.5$) for testing ($N=204$), and found that in the existing samples, the positive correlation region all maintained the original judgment ($N=161$), while in the negative correlation region, 29 were judged negative and 14 were judged positive (i.e., a misjudgment rate of about $0.325$).

Therefore, this method cannot well solve the negative correlation problem and requires further research.