Randomization Can Reduce Both Bias and Variance: A Case Study in Random Forests