Add Qwen2.5_VL Architecture Definition

[qiuxy28]

When merging Qwen2.5-VL models with mergekit, the system fails with:
RuntimeError:Tensor visual.merger.mlp.1.bias required but not present in model Qwen/Qwen2.5-VL-7B-Instruct
The error occurs because visual.merger.mlp.1 is an activation function without trainable parameters, but our architecture definition incorrectly expects weight/bias parameters.

Changes

1. Add Qwen2.5_VL Architecture Definition

Testing

• Successfully merged Qwen2.5-VL models using both linear and TIES methods
• Benchmark results show expected performance on multimodal tasks

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[qiuxy28]

I have read the CLA Document and I hereby sign the CLA