Understanding Difficulty-based Sample Weighting with a Universal Difficulty Measure