Self-supervised Pre-training Reduces Label Permutation Instability of Speech Separation