feat: select on PR-AUC, cross-validate with a custom splitter

[ancongui]

What & why

Two robust model-selection gaps in AutoML, closed with TDD on real data (breast_cancer):

1. PR-AUC is now a selectable CV metric

average_precision (PR-AUC) was already reported on holdout, but it was not in the CV scoring map — so fit(metric="average_precision") silently fell back to selecting the winner on accuracy (scoring_name(...) → default). On imbalanced binary problems, that picks the wrong model.

Now "average_precision" maps to scikit-learn's average_precision scorer (greater-is-better), so the leaderboard, the refit winner, and result.cv_scoring all reflect PR-AUC:

result = AutoML().fit(train, metric="average_precision")   # selected by PR-AUC, not accuracy

2. AutoML(cv=...) accepts a scikit-learn splitter

cv was typed int only, yet it's passed straight to cross_val_score, which already accepts any splitter. Broadened to int | Any and documented + tested the cross-validation strategies that avoid silent leakage:

AutoML(cv=TimeSeriesSplit(n_splits=5))   # temporal: forward-chaining, no future leakage
AutoML(cv=StratifiedKFold(5, shuffle=True))
AutoML(cv=GroupKFold(n_splits=5))        # grouped: keep a group out of train & test

Tests (TDD)

• test_pr_auc_is_a_selectable_cv_metric, test_automl_selects_on_pr_auc — RED→GREEN (asserted 'accuracy' == 'average_precision' before the fix).
• test_automl_accepts_a_cv_splitter, test_automl_accepts_a_time_series_splitter — lock the now-documented cv-splitter contract.
• No regressions: evaluation + automl + ensemble + calibration suites green locally; ruff format/check + pyright(src) clean.

Docs

docs/automl.md: new "cv accepts a splitter" and "select on PR-AUC for imbalanced binary" tips; the binary metric panel now lists average_precision + brier_score.