Custom fast FSM for emphasis correctness (with implementation).

[Yoshqu]

Hi. This always has been a role play immersion breaker for me.

*looks at you "why* this is here?"

Yes, the grammar should help but it's slow. I have implemented FSM which forbids tokens with * when generation is inside of ", and vice-versa.

https://github.com/Yoshqu/koboldcpp-experimental/blob/concedo/experimental/README.md
https://github.com/Yoshqu/koboldcpp-experimental/blob/concedo/experimental/emphasis.h

It is fast, (could be done little bit faster, but i already invested more time that previously anticipated - but hey, it is fun). Feel free to experiment and ask any questions.

It works for me quite well. I could reformat/bring improvements for main koboldcpp repository if it would benefit others...