[Feature request] Support training for new MLLM: Bee-8B

Describe the feature
We have recently released a new Multimodal LLM named Bee-8B.

It is worth mentioning that our associated dataset, Honey-Data-15M, has already garnered significant attention and traction on the ModelScope community.

Given the popularity of the data and the model, many users are looking for a reliable way to fine-tune this model. I would like to request official support for training and inference of Bee-8B in ms-swift. Since ms-swift is the go-to framework for the ModelScope ecosystem, having support would greatly benefit the community.

Model Information:

• Model Name: Bee-8B
• Hugging Face Repo: https://huggingface.co/Open-Bee/Bee-8B-RL
• ModelScope Repo: https://modelscope.cn/models/Open-Bee/Bee-8B-RL
• Home: https://open-bee.github.io/
• Paper: https://arxiv.org/pdf/2510.13795
• Architecture: It is based on Qwen3-8B as the LLM backbone and SigLIP2 as the vision tower.

The issue names no files, tests, or entry points. Review the Bee-8B Hugging Face and ModelScope repositories, home page, and paper to determine the required training and inference behavior, then compare it with existing ms-swift multimodal support. Done means official support is added and validated, but no specific tests or acceptance criteria are provided.