Learning Adversarially Fair and Transferable Representations