On the Implicit Adversariality of Catastrophic Forgetting in Deep Continual Learning