Near-Optimal Bounds for Learning Gaussian Halfspaces with Random Classification Noise