Learning Halfspaces with the Zero-One Loss: Time-Accuracy Tradeoffs