Unlocking Accuracy and Fairness in Differentially Private Image Classification