Constant and near-constant features in actg, physionet_challenge_2012 and diabetes #140

Problem

After dataset generation, these features hold one value for every instance, or almost every one:

• actg feature zprior (ZDV before the trial): 1 for all 2,139 patients, so 0 after z-normalization.
• physionet_challenge_2012 feature MechVent: the Challenge records it only as 1, so a stay's mean is 1 or missing, and mean-filling makes it 1 everywhere. Whether ventilation was recorded is informative (16.6% vs 10.2% in-hospital deaths), but the fill discards it (Mean imputation in TabularCSVDataset discards informative missingness and averages category codes #142).
• diabetes columns 40 and 41 (one-hot smoking categories): 1 for 18 and 40 of 92,062 instances.

A selection of such a feature costs budget and reveals nothing (or, for the rare one-hots, almost nothing). Methods may learn to avoid them, but they inflate the feature count and make budgets less comparable across datasets.

Scope

• Dataset generation for actg, physionet_challenge_2012 and diabetes; both budget settings.
• Changes paper results if features are dropped: actg and physionet_challenge_2012 are in the kdd26 dataset list.

Options to decide

• Drop constant features (zprior, MechVent as generated today).
• For MechVent, keep it as a "was ventilation recorded" indicator instead, if missingness is handled differently.
• For the rare diabetes one-hots, merge them or drop them; depends on Rebuild diabetes from NHANES, and decide what to do about its imputed labels #137, the diabetes rebuild.

Acceptance criteria

• No dataset has a constant feature, or its page says why it keeps one.
• just qa passes.