Noise-Aware Differentially Private Regression via Meta-Learning