Learning under Label Noise through Few-Shot Human-in-the-Loop Refinement