Fix corpus reference orientation for chrF/chrF++/TER metrics

[dyurchenko98]

Summary

Fixes incorrect reference layout for corpus chrf, chrf++, and ter in CorpusLevelTranslationMetric.compute_corpus.

SacreBLEU expects references shaped as [ref_id][sample_id], but lighteval passed [sample_id][ref_id].
This can drop hypotheses from scoring and compare remaining hypotheses against mis-grouped (pooled) references.

Example (single-ref per sample):

• preds: [p1, p2]
• wrong refs passed: [[r1], [r2]]
• interpreted as 2 reference streams of length 1, so only p1 is scored (against refs [r1, r2]), while p2 is ignored.

Changes

• Transpose non-BLEU references with zip_longest(..., fillvalue=None) before calling sacrebleu.
• Keep BLEU path unchanged.
• Add regression tests for:
  • all hypotheses being scored,
  • variable number of references per sample.
• Update affected corpus metric fixture expected values (chrf, chrf_plus, ter) to match correct behavior.

Fixes #1112.