DEAL: Data-Efficient Active Learning for Regression Under Drift