Differentially Private and Fair Deep Learning: A Lagrangian Dual Approach