Private Fine-tuning of Large Language Models with Zeroth-order Optimization