Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference