Bias-Aware Loss for Training Image and Speech Quality Prediction Models from Multiple Datasets