Realized that the logic in #23814 was more than needed, and harder to
maintain. Something like that could make sense if using the tokenizer
and wanting to precisely hit a token limit. However in the case of edit
predictions it's more of a latency+expense vs capability tradeoff, and
so such precision is unnecessary.

Happily this change didn't require much extra work, just copy-modifying
parts of that change was sufficient.

Release Notes:

- N/A