Rethinking Clustering-Based Pseudo-Labeling for Unsupervised Meta-Learning