Refactor witness interpreter o1vm

[marcbeunardeau88]

Currently the witness interpreter state is partially reset at each vm step to provide one line of the witness matrix at a time
This lead to unnatural things such as : some part of the state are permananent, other not, we build a matrix line by line in the main, we allocate one by one, the batch inversion is done separately etc...

I'd like to propose a new api where a witness interpreter env allocate the whole matrix and fill it
Then some pieces of code such as the construction of prrof input, the lookup state padding, the batch inversion could be done here, and not scattered all over the place

In the same PR we cna change the lookup state to be array, not vec, to be consistent with the rest of the state