On the Interaction Between Differential Privacy and Gradient Compression in Deep Learning