High-Quality Prediction Intervals for Deep Learning: A Distribution-Free, Ensembled Approach