On Optimal Hyperparameters for Differentially Private Deep Transfer Learning