Support binary classification in conformal prediction set methods

[fbonc]

Problem
Every SetPredictor subclass only accepted multiclass base models. A binary model emits a single positive-class probability y_prob of shape (N, 1), but a conformal prediction set ranges over both labels {0, 1} and needs one column per class, so binary tasks were out entirely, and the binary metrics function couldn't score prediction sets at all.

Fix
Re-present a binary model as a 2-class problem at the two boundaries where a raw (N, 1) probability enters the pipeline: expand it to [P(y=0), P(y=1)] (N, 2) before the non-conformity scoring in calibrate (for the threshold) and in forward (for the set), so the existing split-conformal machinery runs unchanged with K = 2. This sits on top of the new scores.py scoring layer, so binary composes with any score_type (threshold or APS). Then wire the prediction-set metrics into binary_metrics_fn so results are scorable.

Changes

• Modified pyhealth/metrics/binary.py: binary_metrics_fn now accepts y_predset and computes the conformal set metrics.
• Modified pyhealth/metrics/prediction_set.py: added PREDICTION_SET_METRICS and compute_prediction_set_metric, a single dispatch point for set metrics.
• Modified pyhealth/metrics/multiclass.py: delegates set-metric dispatch to the shared helper (removes duplicated block).
• Modified pyhealth/calib/utils.py: added binary_to_2col (primitive) plus expand_binary_cal (calibrate side, numpy) and expand_binary_pred (forward side, numpy) which expand the probability before the scores.py score functions.
• Modified pyhealth/calib/predictionset/{label,base_conformal,cluster/cluster_label,cluster/neighborhood_label,covariate/covariate_label}.py: accept binary in the mode gate, and expand y_prob to 2 columns before the true_class_* / all_class_* score functions in calibrate and forward.
• Modified pyhealth/calib/predictionset/scrib/__init__.py: accept binary in the mode gate; SCRIB keeps its own quicksearch scoring, so the expansion is an inline torch step (it thresholds the probability directly rather than going through scores.py).
• Created tests/core/test_binary_label.py, tests/core/test_binary_predictionset.py: binary end-to-end coverage for all six methods and the helpers.
• Modified tests/core/test_{cluster,covariate,neighborhood}_label.py: the "binary raises" tests now assert binary is accepted and that a genuinely unsupported mode still raises.

Core decisions

• y_prob stays native (N, 1) - only the set is expanded. roc_auc, etc. are unaffected, and self.mode stays honest ("binary") rather than being faked as multiclass.
• Expansion happens in both calibrate and forward. They're different data; only the threshold self.t passes between them, so each entry point must expand its own raw probability.
• Binary composes with the scores.py layer: expanding to (N, 2) before the score functions means score_type (threshold / APS) works for binary with no extra code.
• Set-metric dispatch is centralized in prediction_set.py, and the binary and multiclass metrics functions delegate to the shared helper (previously duplicated in multiclass.py).
• FavMac is intentionally excluded. It is multilabel, and a binary task doesn't map to its cost-over-multiple-labels contract.