docs: add normalized row example and failure cases to dataset loader docs

[CalebWang0126]

Summary

Refs #42. The dataset loader tutorial landed earlier, but two of the issue's acceptance criteria are still unmet in docs/cli/dataset_loaders.rst:

• "Show a small JSONL or Hugging Face dataset example."
• "Includes failure cases for missing prompt/answer/reward fields."

This PR closes that gap. Docs-only; no CLI changes.

Changes (docs/cli/dataset_loaders.rst)

• Normalized row example: a two-row canonical JSONL snippet (prompt + solutions) that needs no loader, plus how GSM8K-style Hugging Face rows reach the same shape through examples/math/dataset_loader.py, and how solutions flows into the reward as record.answer.
• Common failure cases: the exact startup errors, verified against current code:
  • missing prompt in a rollout row -> dataset row must contain \prompt`; use --dataset-loader-fn to normalize raw rows (areno/api/trainer.py`)
  • missing prompt/response in SFT rows -> SFT dataset loader must return rows with \prompt` and `response`...; None/empty rows are skipped and an all-skipped dataset fails with SFT dataset produced no valid training rows after filtering (areno/api/trainers/sft.py`)
  • missing chosen/rejected in DPO rows -> DPO dataset row must contain \chosen` and `rejected` (areno/api/trainers/dpo.py`)
  • missing solutions for the math reward -> math reward expects \record.answer`... (examples/math/math_verify_reward.py`)

Verification

• Error strings quoted verbatim from the current implementation.
• sphinx-build -b html -E docs builds with zero warnings.
• Uses the existing math examples as source references, per the acceptance criteria.