rewrite numpy features creation to be effiecient in the cv loop

[furlat]

create categorical encodings only once, possibly in polars.
create numpy arrays only once or try at least to do it memoryless in polars with a single filter one the validation
consider trade off between computation and memory by precaching features before modal deployment