Efficiently Learning Adversarially Robust Halfspaces with Noise