Exploring the Limits of Differentially Private Deep Learning with Group-wise Clipping