Ensemble Performance Through the Lens of Linear Independence of Classifier Votes in Data Streams