[ChrisW09]

Fixes #493

Problem

SklearnBaseClassifier.evaluate() passed the original y_true labels to the probability-based metrics, together with predict_proba() output whose columns follow classes_. score(X, y, metric=(metric, True)) did the same. Two failures followed:

• For string labels, AUPRC, BrierScore and ExpectedCalibrationError failed internally and were silently reported as nan.
• For integer labels other than 0..K-1 (e.g. {1, 2}), AUPRC scored the wrong class (0.302 instead of 1.0) and ECE became nan.

Fix

• A new helper, _probability_metric_targets, encodes y_true with the existing _encode_labels(…, classes_). That is the same 0..K-1 mapping fit() and predict() use, and it matches the predict_proba() column order.
• evaluate() passes the encoded labels to the probability-based metrics (auroc, auprc, log_loss, brier, ece). Label metrics (accuracy, F1, …) still receive the original labels next to the decoded predict() output.
• score(X, y, metric=(metric, True)) does the same for those DeepTab probability metrics. Other callables keep receiving the original labels, so user-supplied functions see the labels they were given.
• Labels that cannot be encoded (a label never seen during fit, or a None) fall back to the raw labels. The existing per-metric error handling then applies, so evaluate() itself doesn't start raising where it didn't before.

The names of the probability metrics move from a local set in evaluate() to the module-level _PROBA_NAMES, so both methods use the same list.

#518 (LogLoss with missing classes) treats valid integer labels as column indices and relies on evaluate()/score() passing encoded labels, so this PR should be merged before it.

Tests

tests/integration/test_label_encoding.py::TestClassifierLabelEncoding:

• test_probability_metrics_receive_encoded_labels, for string binary, {1, 2}, {10, 20} and string multiclass labels: every probability metric is finite, and both evaluate() and score(..., metric=(metric, True)) equal the same metric computed directly on the encoded labels and predict_proba(). Binary-only metrics are used only in the binary cases.
• test_evaluate_reports_nan_for_labels_that_cannot_be_encoded: a None in y_true gives nan, as on main, instead of a TypeError.

The 4 parametrized cases fail on main and pass with this change. The None case guards the fallback.

Docs

The evaluate() and score() numpydoc say that DeepTab's probability metrics receive labels encoded to the classes_ column indices, and everything else the original labels.

Checks

• just test: 3241 passed, 46 skipped, 28 xfailed; coverage 91.37%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)