Stochastic Engrams for Efficient Continual Learning with Binarized Neural Networks