In config generation, switch to two stage training when the mono data is too small

[gregtatum]

From: #620 (comment)

But I found another use case when we don't want ["one-stage" teacher training when using a pre-trained backtranslations model]. If the amount of mono-trg data is too small (for example for en-lt) we still want to use two-stage. We don't want to loop over 5M back-translated sentences.

[eu9ene]

Now with HPLT2, NLLB and Monocleaner we always have a lot of mono data. I'd say we still use two stage by default and switch to one stage if it stops too early.