Fix CorpusLevelF1Score(None) to report positive-class F1, not max-per-class

[iamsharduld]

What

CorpusLevelF1Score.compute_corpus (the num_classes == 2 path) returns
np.max(f1_score(golds, preds, average=self.average)). When average=None — which is how
Metrics.loglikelihood_f1 instantiates it (CorpusLevelF1Score(None), used by glue:mrpc and
glue:qqp) — f1_score returns the per-class F1 array, so np.max reports the best class's
F1 rather than the positive-class binary F1, inflating the score.

golds = [0,0,0,0,0,0,1,1,1,1]
preds = [0,0,0,0,0,1,0,0,0,1]
# per-class F1: [0.714 (class 0), 0.333 (class 1)]
CorpusLevelF1Score(None).compute_corpus(items)  # 0.714  (np.max)  -- should be 0.333

Fix

When average is None, return the positive class fscore[1]. Scalar averages
(micro/macro/weighted) are unchanged (np.max of a scalar was a no-op).

Tests

tests/test_unit_base_metrics.py::test_corpus_level_f1_binary_positive_class asserts the
positive-class F1 (1/3) on a fixture where the max-per-class would be ~0.714. Fails before,
passes after.