Maintaining Plasticity in Continual Learning via Regenerative Regularization