Fix ECE confidence binary crash

[fbonc]

Issue

ece_confidence_binary indexed prob[:,0] and label[:,0] (calibration.py:150), requiring 2D arrays. Its only real caller, binary_metrics_fn (binary.py:91), passes 1D positive-class probabilities and 1D 0/1 labels, so ECE and ECE_adapt always raised IndexError for binary tasks via the documented API. As a secondary defect, even for 2D input the code used column 0 (class 0) despite the docstring stating the metric is computed on class 1.

Fix

Compute confidence from the positive class (prob[:,1] when 2D, else the 1D array) and use the 0/1 label as the accuracy target (label[:,1] when 2D one-hot, else the 1D array). This makes the function work with the 1D shapes the caller actually uses and corrects the class-0/class-1 indexing. ece_classwise is unaffected (it calls a separate internal helper, not this function).

Notes

Added new test file tests/core/test_calibration_binary_ece.py with two tests. No prior tests existed for this function.

[DarylOkeke]

I think this still crashes with PyHealth's normal binary model output. BinaryLabelProcessor.process() returns shape (1,), so after batching and inference both prob and label are (N, 1). Since ndim > 1, this code tries [:, 1], but column 1 does not exist. Could the shape handling distinguish flat (N,), single-column (N, 1) (use column 0), and two-column (N, 2) (use positive-class column 1), with a regression test for (N, 1)?

[fbonc]

I think this still crashes with PyHealth's normal binary model output. BinaryLabelProcessor.process() returns shape (1,), so after batching and inference both prob and label are (N, 1). Since ndim > 1, this code tries [:, 1], but column 1 does not exist. Could the shape handling distinguish flat (N,), single-column (N, 1) (use column 0), and two-column (N, 2) (use positive-class column 1), with a regression test for (N, 1)?

fixed!