Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals