Noisy Correspondence Learning with Self-Reinforcing Errors Mitigation