Flexibly Fair Representation Learning by Disentanglement