Tight Lower Bounds on Worst-Case Guarantees for Zero-Shot Learning with Attributes