Pre-training Differentially Private Models with Limited Public Data