add: unified_reward and qwen_image_bench

[yjy415]

No description provided.

[gemini-code-assist]

Code Review

This pull request adds support for three new image quality metrics: Qwen-Image-Bench, UnifiedReward 2.0, and UnifiedReward Edit, including their respective models, metrics, state dict converters, documentation, and example scripts. The code review identified several critical issues: an accidental deletion of hidream_o1_image_series from the model configurations, a potential ValueError crash when parsing pairwise rank winners in UnifiedRewardEditModel, a potential AttributeError in fix_score_json if the input is not a dictionary, a greedy regex pattern that could break JSON extraction, and a redundant regex character class [::]. Additionally, the input normalization logic in UnifiedRewardEditModel can be simplified.