Sample Selection via Contrastive Fragmentation for Noisy Label Regression