Playing NetHack with LLMs: Potential & Limitations as Zero-Shot Agents