fix: transpose references before passing to sacrebleu in CorpusLevelTranslationMetric

[jaydenC88]

Summary

Fixes #1112

CorpusLevelTranslationMetric.compute_corpus() was passing references to sacrebleu in the wrong shape. References were collected as [sent_id][ref_id], but sacrebleu's corpus_score expects [ref_id][sent_id]. Because sacrebleu internally does zip(*references) to transpose the streams, the wrong orientation meant only the first hypothesis was ever scored — matched against all references from all sentences pooled together — causing inflated scores (e.g. 100.0 for a half-correct corpus).

Changes:

• src/lighteval/metrics/metrics_corpus.py: add from itertools import zip_longest and replace the per-metric special-case logic with a single zip_longest(*golds) transpose applied uniformly before corpus_score(). This fixes chrF, chrF++, and TER, and also lets BLEU use all available references instead of silently dropping to gold[0].
• tests/unit/metrics/test_cases/chrf.json, chrf_plus.json, ter.json: update expected values that were computed with the buggy code.

Reproducer (from issue)

from lighteval.metrics.metrics_corpus import CorpusLevelTranslationMetric
from lighteval.metrics.sample_preparator import GenerativeCorpusMetricInput
from lighteval.utils.utils import as_list

items = [
    GenerativeCorpusMetricInput(golds=["GOOD"], preds=["GOOD"]),
    GenerativeCorpusMetricInput(golds=["REF2"], preds=["PRED2"]),
]
metric = CorpusLevelTranslationMetric("chrf++")

# Before fix: stats=1, score=100.0 (only first hyp scored against all refs)
# After fix:  stats=2, score=58.7  (each hyp scored against its own ref)
print(metric.compute_corpus(items))

Test plan

• Reproducer from issue confirms len(stats) == len(hypotheses) after fix
• All 8 chrf / chrf++ / ter / bleu test cases pass (pytest tests/unit/metrics/test_automated_metrics_pytest.py -k "chrf or ter or bleu")