main: port basic LLaVA (multimodal) support from llava-cli

[Nekotekina]

Allows interactive chat in the terminal and possibly other features I haven't tested.
Example:

~/github/llama.cpp/build/bin/main\
-m ~/Downloads/LLM-Models/llava-v1.6-34b.Q5_K_M.gguf\
--mmproj ~/Downloads/LLM-Models/mmproj-llava-34b-f16-q6_k.gguf\
--image <(convert -resize 448x448\> <(xclip -selection clipboard -t image/png -o) PNG:-)\
-tb $(nproc) -c 4096 -n 1024 -ngl 10 --temp 0.2 --repeat-last-n 1024 -s 0\
--color -cml -p "Give detailed answers to user's question about the image: <image>"

As in llava-cli, <image> gets replaced with actual image embedding. Not sure about 448x448 size, this may be wrong, but this is just an example of image preprocessing (X server clipboard -> ImageMagick).

[Nekotekina]

If this is unwelcomed you could just close it...

[lin72h]

Wow RPCS3's developer is helping us! Thanks

[slaren]

If this is unwelcomed you could just close it...

It's not unwelcomed, but nobody is working on the llava code at the moment, so it is hard to review. The llava code is also pending a refactor so that it can be added back to the server (#4216 #6027). After that it may make more sense to add this functionality to other examples.