Reject corpus length mismatches when using cached references

[sueun-dev]

corpus_score(hyps, None) silently truncates to the shorter of the hypotheses and the cached references. With one cached reference, a matching first hypothesis and an extra incorrect hypothesis still produce a perfect BLEU/chrF score or zero TER. Passing those references explicitly already raises a length error after #294; this covers the cached path left over from that validation work (related to #220 and #281).

Check the selected reference cache's length before extracting statistics. Putting the check here also covers PairedTest, which calls the extractor directly. Tests cover shorter and longer hypothesis lists across BLEU/chrF/TER, valid cached scores with missing alternative references, explicit-reference overrides, sentence scoring, and mismatched baseline/candidate inputs in serial and multiprocessing significance tests.

Checked on Python 3.12.14, Linux arm64:

• Final regression file: 14 failures on the original source; 64 passes with the fix.
• python -m pytest -q --ignore=test/test_dataset.py -k 'not test_api_get_source and not test_api_get_reference': 197 passed, 6 deselected.
• mypy sacrebleu scripts test and Ruff on all tracked Python files passed.
• Source distribution and wheel built offline with synthetic 2.6.1.dev0 version metadata.

The dataset-download tests, CLI shell suite, optional Japanese/Korean tokenizers, and other Python/OS combinations were not run. An initial Ruff run also scanned generated build/version files in the source archive; the tracked-source check follows the repository's ignore rules.