On Zero-Shot Counterspeech Generation by LLMs