A Split-and-Privatize Framework for Large Language Model Fine-Tuning