Learning Fair Representations via Rate-Distortion Maximization