CarLLaVA: Vision language models for camera-only closed-loop driving