End-to-end Gemma 4 setup on mainline llama.cpp with text and vision throughput notes.