Filling the gap between GRACE-and GRACE-FO-derived terrestrial water storage anomalies with Bayesian convolutional neural networks