build llama.cpp inside a Docker container with AMD ROCm support