[Feature] Add support for pointwise reranker training without pairwise data

[Tord-Zhang]

Checklist / 检查清单

• I have searched existing issues, and this is a new feature request. / 我已经搜索过现有的 issues，确认这是一个新的 Feature Request。

Feature Request Description / Feature Request 描述

After carefully review the code, I am sure right now the reranker training requires both positive_messages and negative_messages in each training sample. So would you please add support for point-wise reranker training without pair-wise data? Since in many scenarios, pairwise data are unreachable.

Pull Request / Pull Request 信息

No response