MaxMatch: Semi-Supervised Learning With Worst-Case Consistency