Fix macro average calculation in geometric_mean_score for binary classification

[shubhrai23]

Geometric_mean_score with average='macro' was incorrect for binary classification (it was calculating the arithmetic mean of scores instead of the geometric mean of sensitivity and specificity).

Added a check to handle binary cases correctly while preserving multiclass behavior.

Added a regression test.

[immu4989]

Thanks for tackling #1096 — the diagnosis (binary macro returning sqrt(mean_tpr * mean_tnr) instead of mean(sqrt(tpr*tnr))) is correct. But I think the fix is incomplete, and I wanted to flag it constructively before it lands.

The bug isn't binary-specific — multiclass macro is wrong too. This PR only switches to the per-class computation when is_binary is true, so the multiclass path still returns the old (incorrect) value. On the reporter's own multiclass fixture:

y_true = [0, 1, 2, 0, 1, 2]; y_pred = [0, 2, 1, 0, 1, 2]
geometric_mean_score(y_true, y_pred, average=None)     # [1. , 0.612, 0.612]
np.mean(...)                                            # 0.7416  <- expected
geometric_mean_score(y_true, y_pred, average="macro")  # 0.7454  <- still wrong with this PR

I verified this by checking out the branch locally — after the patch, macro still ≠ np.mean(average=None) for multiclass.

Related concern: the existing parametrized tests (test_geometric_mean_average, test_geometric_mean_score_prediction) still assert the old macro/weighted numbers, and they pass here only because the multiclass path is unchanged. So those expected values are themselves part of the bug and should have been updated — the fact that they still pass is a signal the multiclass case wasn't addressed.

The root cause is the same for binary and multiclass (sqrt of the averaged sen/spe vs. the average of per-class G-means), so it can be fixed unconditionally without the is_binary branch. PR #1173 takes that approach — computes per-class G-means and aggregates for both macro and weighted, updates the now-incorrect expected values, and adds an invariant test that macro == mean(per_class). It might be cleanest to consolidate on that one, or to drop the is_binary gate here and refresh the multiclass expectations.

Either way, removing the special-casing and fixing the multiclass expected values would make this complete.