Add cosine_similarity to hn_mine

[daegonYu]

By specifying a range for similarity scores when mining hard negatives, this argument ensures that negative examples fall within a desired difficulty level. This fine-tuned control helps in avoiding extremes—negatives that are either too close or too far in meaning from the query.

This is also explained in the paper (https://arxiv.org/pdf/2405.05374 (Appendix Algorithm 1: Tunable Negative Mining)), and I also used this code to mine hard negatives, and as a result, I was able to create a Reranker model that performed better than the hard negatives mined with the existing code. I would like to contribute to others using this code to create good models.