Use latest UniMERNet model for Formula Recognition

[ChenZiHong-Gavin]

I noticed that the UniMERNet model weight provided is not the latest version. The latest version is larger but better, and can lead to better formula recognition results, as I have also observed many complaints about inaccurate formula recognition in the issues.

For example:

• 公式检测和公式识别不准确 MinerU#537
• 数学符号会识别成特殊字符 MinerU#280
• 公式识别异常 #20

latest UniMERNet version:
huggingface: https://huggingface.co/wanderkid/unimernet/tree/main
modelscope: https://www.modelscope.cn/models/wanderkid/UniMERNet/files

test img:

former model leads to confusing result:

latest model will output the right answer:

[ChenZiHong-Gavin]

noticed you have updated unimernet_base | unimernet_small | unimernet_tiny
however, the unimernet version is still not the latest.
your pytorch_model.bin is 3.75 GB

in https://huggingface.co/wanderkid/unimernet/tree/main
the pytorch_model.bin is 4.91 GB