SOLA: Continual Learning with Second-Order Loss Approximation