Beyond Random Sampling: Instance Quality-Based Data Partitioning via Item Response Theory