Denoising Low-Rank Data Under Distribution Shift: Double Descent and Data Augmentation