fix(metrics): count confidence 1.0 in ECE and stop thresholding 1-D integer labels

[ChrisW09]

Fixes #421

Problems

1. ExpectedCalibrationError ignores confidence == 1.0. Every bin was half-open [lo, hi), including the last [0.9, 1.0), so predictions at exactly 1.0 fell into no bin and contributed 0. A model predicting the wrong class with probability 1.0 on every sample — the worst possible calibration — scored ECE = 0.0.
2. Accuracy/F1Score binarize 1-D input at 0.5. The docstrings promise 1-D integer-label support, but any 1-D input was thresholded, mangling multiclass labels (label 2 → 1). Accuracy()([0,1,2,2], [0,1,2,2]) returned 0.5 for perfect predictions. This bites the natural composition Accuracy()(y, model.predict(X)), since predict returns 1-D labels.

Fix

• Make the final ECE bin right-inclusive.
• Add a shared _labels_from_predictions helper: 2-D input → argmax, 1-D integer-valued input → labels as-is, 1-D fractional input → thresholded at 0.5 (so binary probability scores keep working).

Tests

test_ece_counts_full_confidence, test_accuracy_1d_multiclass_labels, test_f1_1d_multiclass_labels. Full tests/test_metrics.py passes.

🤖 Generated with Claude Code

[ChrisW09]

CI note: the two red test jobs on this PR (test_experimental_lss_fit_predict_evaluate[TromptLSS] on py3.11/windows and test_experimental_classifier_fit_predict_evaluate[TromptClassifier] on py3.12/ubuntu) are unrelated to this change — this PR only touches deeptab/metrics/classification.py and its tests.

They are a pre-existing flake: Trompt.init_rec is created with nn.Parameter(torch.empty(...)) and never initialized, so it can start as NaN depending on what is in the allocation (issue #423, item 6). Fixed in #440, which reproduces it deterministically.