Differentially Private Continual Learning using Pre-Trained Models