Trying out llama.cpp’s new vision support