Towards Adversarial Evaluations for Inexact Machine Unlearning