Optimal Unbiased Randomizers for Regression with Label Differential Privacy