Continual Learning by Asymmetric Loss Approximation with Single-Side Overestimation