Differentiable Optimal Adversaries for Learning Fair Representations