[Feature] Vision projector for embedding models

[MerijnHendriks]

Describe the Issue

I want to use the vision capabilities from qwen-3-vl-embedding for embedding images/ocr, but there is no place in koboldcpp where I can load the embedding model's mmproj.

Request

Add a place to load the embedding model's mmproj in the interface.

Temporary workaround

Using llama.cpp, start a new server for both the embedding model + mmproj, and enable embeddings flag.
Then use that server for embedding instead of koboldcpp.

For testing, you can use:

• https://huggingface.co/DevQuasar/Qwen.Qwen3-VL-Embedding-2B-GGUF/resolve/main/Qwen.Qwen3-VL-Embedding-2B.f16.gguf
• https://huggingface.co/DevQuasar/Qwen.Qwen3-VL-Embedding-2B-GGUF/resolve/main/mmproj-Qwen.Qwen3-VL-Embedding-2B.f16.gguf

[LostRuins]

I think for now this is a little out of scope for koboldcpp as piping visual input to the embedding model is not easy to do in the current setup. We can revisit this idea in future.