Multi-Domain Long-Tailed Learning by Augmenting Disentangled Representations