How Ensemble Learning Balances Accuracy and Overfitting: A Bias-Variance Perspective on Tabular Data