Vision support (e.g. gemma3/llava)

[getnamo]

With ggml-org/llama.cpp#12344 merged, it should be possible to support e.g. gemma3 vision capabilities. Wishlisty for now until we get more dev time, but potentially compelling for e.g. inferencing something shown in a game at runtime (whether a PoV or sketches/etc).

Resources:

• ggml - #12348 - instructions to build and run llama-gemma3-cli
• llama.cpp - gemma3-cli.cpp code that is used to run image embedding.

Todo:

• Rough implementation showing 1:1 image inference for a fixed target
• Streamline input to be properly queued (threaded)
• InsertImage API
• InsertTemplatedPromptWithImage combined call API
• Auto-convert file UTexture2D* type to image understood by clip encoder
• Test Gemma3 4B and 12B capabilities running in engine
• Auto-convert any image file format into UTexture2D* and embed (for e.g. network receive or file read)

[getnamo]

NB:

• updated pull req for vision support here: server : vision support via libmtmd ggml-org/llama.cpp#12898
• docs for mtmd-cli: https://github.com/ggml-org/llama.cpp/blob/master/docs/multimodal.md
• mtmd location: https://github.com/ggml-org/llama.cpp/tree/master/tools/mtmd
• mtmd-cli code: https://github.com/ggml-org/llama.cpp/blob/master/tools/mtmd/mtmd-cli.cpp
• header for api: https://github.com/ggml-org/llama.cpp/blob/master/tools/mtmd/mtmd.h