An Optimization Framework for Differentially Private Sparse Fine-Tuning