Meta Self-Refinement for Robust Learning with Weak Supervision