Beyond Perturbations: Learning Guarantees with Arbitrary Adversarial Test Examples