Handle small-sample cross-fitting folds gracefully

[TomeHirata]

Follow-up to the JOSS review's minor point on empty folds (openjournals/joss-reviews#11047):

#150 added a check for one case only: a fold leaving no target-arm training data. Most small-sample failures took other paths and still surfaced as opaque errors.

Changes (dte_adj/stratified.py, dte_adj/util.py)

• Empty prediction folds are skipped. A fold (or a fold × stratum cell) with no records used to call model.predict_proba on an empty array. sklearn then raised Found array with 0 sample(s). Such a cell has nothing to predict, so it is now skipped. This was the most common failure, and those runs now simply succeed.
• Informative errors for genuinely untrainable cells. The upfront check now also covers:
  • a fold × stratum cell with no target-arm training observations. This used to raise IndexError: index 0 is out of bounds. The error now names the fold and stratum and suggests reducing folds or merging small strata.
  • a stratum with no observations of the target arm at all.
• Multi-task bug fix. When a fold × stratum cell's training labels were all identical, no model was fit for it. The code then reused the model from the previous cell, which was fit on another fold or stratum. If no model existed yet, it raised AttributeError: ... has no attribute 'model'. It now predicts the constant label, as the single-task path already does.

Effect (200 random seeds each, LogisticRegression, single-task)

Setting	Before	After
No strata, n=30, 10 folds	119 opaque errors	0 errors
No strata, n=20, 5 folds	25 opaque + 1 informative	1 informative
3 strata, n=20, 5 folds	200 opaque (IndexError / sklearn)	200 informative

With multi-task (LinearRegression), 3 strata, n=20 and 5 folds, 24 runs used to fail with AttributeError and 134 with the informative error. After the change, all 146 failures are informative. Failures can still happen at sizes this small. The difference is that each one now tells the user what went wrong and what to do.

Testing

• New tests: four in TestSmallSampleFolds, covering an empty prediction fold, a stratum without training data, a stratum without the target arm, and multi-task constant labels. All four fail on main and pass with this change.
• Full suite: 83 passed. ruff check . is clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.