On Unsupervised Uncertainty-Driven Speech Pseudo-Label Filtering and Model Calibration