Tempered Sigmoid Activations for Deep Learning with Differential Privacy