Guiding LLMs to Generate High-Fidelity and High-Quality Counterfactual Explanations for Text Classification