Hitting the Target: Stopping Active Learning at the Cost-Based Optimum