Training Differentially Private Ad Prediction Models with Semi-Sensitive Features