Add sampling

[marksverdhei]

We should try out an alternative for bert-insert,
where we sample from the softmax of the logits instead of picking the highest every time.
This can lead to some interesting results. I suggest top-k sampling to not make it too nonsense-prone either

[LBlend]

I support this!