Interest check: HPO-tunable preprocessing as a first-class concept in TabArena?

Quick question, not a proposal yet.

RamanBench (built on `tabarena`/`bencheval`, see #495/#506) tunes its domain-specific preprocessing (baseline correction, denoising, etc.) jointly with each model's own hyperparameters, via a mixin that adds `prep_*` params to the model's search space, so HPO can pick a preprocessing recipe per model per fold rather than applying one fixed pipeline upfront.

TabArena's own preprocessing (`TabArenaModelAgnosticPreprocessing`/`TabArenaModelSpecificPreprocessing`) runs once via the feature-generator path, outside HPO.

Would a general concept of HPO-tunable preprocessing (as part of a model's search space, not a fixed pre-pass) be of interest for TabArena itself, beyond domain-specific downstream benchmarks? Just gauging interest before writing anything up.