Evaluating the Robustness of Deep Learning Models on Automated Program Repair