Other than being in Rust, how does this compare to a CPU-only build of llama.cpp, which is both easy and already well-supported?