Processing cleanups v0

[RBendias]

Cross-cutting

• tests withCUDA across the whole test/processing suite
• reconsider whether input is a good name (shadows the Python builtin) - Don't use input as var name #274
• reorganize package
• why do we need buffers and should they be in v0?
• define nan strategy for all processing → global strategy (including documenting)
• check for necessary early validation
• batch support
• Align processor docs - come up with processor docs rule
• Check that repr is aligned across processors
• Set generator within base model
• Rename processors (SoftmaxTemperatur -> Softmax, potentially ShuffleCategories instead of CategoryShuffle)
• Check any processor code and align with Agents.md
• Add target inverse as processor
• Update Recipe instruction

Power

• Needs to be fixed: run on regression label. Entered **only when the inverse Yeo-Johnson produces ±inf which is the case when model outputs inf which is unrealistic -> done in quantile PR
• Add AGENTS.md #6 — vectorize: replace the per-column transform/inverse loops with a vectorized torch.where form (validated bit-identical) for torch.compile-friendliness including torch.compile smoke test

Quantile (added from PR #96 review)

• revisit the n_quantiles cap int(subsample * 0.2) (reviewer-flagged; align with sklearn min(n_samples, subsample) or document)
• memory-efficient deterministic subsampling — avoid materializing randperm(n_samples)
• include Quantile in the vectorization / torch.compile work (per-column _transform_col loop)

TabICL processors (PR #106 / #107 review)

• SigmaClip: guard NaN std (clamp to min_std, or fall back to stage-1 stats / ±inf bounds) so all-NaN or all-outlier columns don't corrupt forward; add an all-NaN SigmaClip test; document the soft-bound NaN policy (inherited from TabICL OutlierRemover)
• SoftmaxTemperature: reject non-finite temperature (nan/inf), not just <= 0
• MeanImpute: pre-mask ±inf before nanmean so infinities don't leak into fitted means