XLDA: Linear Discriminant Analysis for Scaling Continual Learning to Extreme Classification at the Edge