Add support for processing only explicitly expected categories for preprocessing.OneHotEncoder, preprocessing.OrdinalEncoder, akin to sklearn api for respective encoders.

All doctests pass (i've added some).

Rationale:
sklearn has this neat feature where you can explicitly pass category values you want to see in the encoder state, other values are filtered out. See categories parameter: OneHotEncoder, OrdinalEncoder

This is convenient when you work with high cardinality category spaces where some values are rare and you want to regularize your model. E.g. I've had a practical problem where constraining only to pre-selected top 20% frequent categories in 1 000 000 cardinality space can give you a 10%+ latency boost with no significant loss in metrics, and also make a model lighter on RAM.

This implementation is hackable so if user wants to modify lists of expected categories between training steps, they can do so by direct attribute access. E.g. can glue with modules like TargetAgg for some cool dynamic reevaluation of expected category lists.

P.S. Pls bump Ruff, my LSP config compains coz api changes. Also MyPy complained a lot about about str | dict | defaultdict type hints for category parameter, I just had to give up on them, maybe someone has better ideas how to handle them.