feat(lgbm): support string features via pandas categorical

[aywandji]

add logic to convert string features to int (using lightgbm model pandas categoricals) before prediction

[cp2587]

I let @crqs merge this as he recently worked on this project to make sure the inference performances were good enough: i want him to know of this so he can backtrack if performance deteriorate.

also as discussed, we should see if the python code and loop can be replaces by something else as it has maybe an huge impact on inference times