Revisiting Non-Autoregressive Translation at Scale