Exploiting MMD and Sinkhorn Divergences for Fair and Transferable Representation Learning