fix(metrics): f1_score silently returns 0 when gold/pred is a single article word (#2298)

[nac7]

What

f1_score calls normalize_text() on both gold and pred before computing
token overlap. normalize_text strips English articles (a, an, the)
via re.sub(r"\b(a|an|the)\b", ...).

When the gold (or pred) string is entirely one of those words — for
example, the letter "A" in a multiple-choice benchmark like MMLU —
article removal produces an empty string. set("".split()) == set(), and
nltk.metrics.scores.f_measure(set(), ...) returns None, which
f1_score silently maps to 0.0.

Concrete failure (pre-fix):

>>> f1_score("A", "A")
0.0   # should be 1.0
>>> f1_score("An", "An")
0.0   # should be 1.0

Any benchmark that reports f1_score alongside single-letter gold answers
(MMLU, ARC, HellaSwag …) will show 0.0 for correct predictions. The same
bug applies to the pred side, and to non-English benchmarks where "a" is
a meaningful token rather than an English article.

Fix

src/helm/benchmark/metrics/evaluate_reference_metrics.py

After computing the normalised token sets, if either set is empty, retry
with should_remove_articles=False before calling f_measure. Strings
that contain articles alongside other tokens are unaffected — article
removal fires as before.

# Before (one line):
ret = f_measure(set(normalize_text(gold).split()), set(normalize_text(pred).split()))

# After:
gold_tokens = set(normalize_text(gold).split())
pred_tokens = set(normalize_text(pred).split())
if not gold_tokens:
    gold_tokens = set(normalize_text(gold, should_remove_articles=False).split())
if not pred_tokens:
    pred_tokens = set(normalize_text(pred, should_remove_articles=False).split())
ret = f_measure(gold_tokens, pred_tokens)

Tests

src/helm/benchmark/metrics/test_evaluate_reference_metrics.py — four new
test functions (no prior f1_score coverage existed):

Test	Checks
test_f1_score_normal	Baseline partial-overlap and no-overlap cases
test_f1_score_article_gold	Regression: "A"/"An"/"The" as gold — now scores 1.0 for a correct pred
test_f1_score_article_in_longer_string	Multi-word strings: article removal still fires
test_f1_score_empty_pred	Empty prediction still yields 0

Fixes #2298