Practical Privacy Filters and Odometers with R\'enyi Differential Privacy and Applications to Differentially Private Deep Learning