Re-using Adversarial Mask Discriminators for Test-time Training under Distribution Shifts