Support Regression Tasks

[HanXudong]

Update fairlib to support regression tasks.

To use the original training and evaluation framework for classification tasks, we keep the y label as categorical variables and add an additional variable for the actual target value for regressions.

• We use pandas to discretize regression targets. Specifically, the number of bins is specified as a hyperparameter, and each bin has the same size.
• The regressor is trained to minimize MSE loss, and other settings are exactly the same as the classification model.
• We currently report MAE, MSE, and R2 as the regression performance metrics.
• Debiasing and fairness evaluation are largely based on bin labels, so we can still report metrics like TPR GAP and FNR GAP
• We have also updated FairBatch to support regression debiasing, where loss estimation and resampling probability adjustment are based on MSE loss.