Fix Jev task questions, ordinal typing, and extreme request lengths

[sileod]

Summary

Audit-driven fixes for tasksource-jev-typed-decisions:

• give ambiguous classification families explicit task questions instead of the generic criterion prompt
  • CrowdFlower configs, including economic-news and tweet_global_warming
  • TuringBench, I2D2, args_me, Touche23-ValueEval, starcon, subjectivity
  • MOH, VUAC, TroFi
  • AmbigNQ
• support config-specific question metadata in shared Tasksource mappings
• mark XGLUE WPR as ordinal so its Bad → Fair → Good → Excellent → Perfect scale can emit score examples
• filter pathological sequence lengths on the complete rendered Jev request
  • default 128 KiB UTF-8 cap
  • optional exact token cap when a tokenizer is supplied
  • record per-split dropped-row counts in the build report
  • include the new limits in the shard fingerprint
• add regressions for the question metadata, ordinal flag, build fingerprint, byte cap, and exact-token cap

Audit notes

The NusaX choice/score mixture was investigated and is intentional: ordinal tasks deliberately produce both representations, so it is unchanged.

Legacy mojibake / HTML noise is also unchanged. The Jev recast already has an explicit raw-text robustness path; this PR does not apply blanket text normalization.

No source labels are changed by this PR.