fix: use pre-compiled llama-cpp-python wheel + model in image

fix: bake GGUF model into Docker image during build

fix: use pre-compiled llama-server binary (zero compilation)

init