The Interplay of Statistics and Noisy Optimization: Learning Linear Predictors with Random Data Weights