Opt/encoders/small transform

[cakedev0]

Reference Issues/PRs

See also scikit-learn#32368, which proposes a similar small-batch fast path but specific to TargetEncoder. This PR implements the same idea more generally, at the _check_X/_encode level shared by OneHotEncoder,
OrdinalEncoder and TargetEncoder (they all go through _BaseEncoder._transform).

What does this implement/fix? Explain your changes.

Speeds up repeated small-batch transform calls (e.g. scoring one row at a time in a serving loop) for OneHotEncoder, OrdinalEncoder and TargetEncoder.

To achieve that, I cache the mapping used by _encode for the object arrays path through _map_to_integer (other paths are: numerical arrays & non-numerical Series). This mapping uniques -> index costs O(n_categories) to build, it's what makes this path slow.

Any sized objects arrays will benefit from this optimization.

The Series path has the same problem, so I make _check_X convert small pandas/polars inputs to arrays so it can benefit from this exact same optimization.

The numerical path doesn't suffer from such an O(n_categories) cost, so it's left unchanged.

AI usage disclosure

I don't remember 😅

Benchmarks

TODO: benchmarks to follow.