Step-E: A Differentiable Data Cleaning Framework for Robust Learning with Noisy Labels