Self-Supervised Learning of Audio Representations from Permutations with Differentiable Ranking