Detecting Dataset Drift and Non-IID Sampling via k-Nearest Neighbors