docs: fix reward function API contract drift

[CalebWang0126]

Summary

Fixes #141.

Some documentation still described the reward contract as reward_fn(example, completions) -> list[float], while the implementation loads reward_fn(record) -> float over a RewardRecord (areno/api/rewards.py). This PR retires the old batch shape everywhere and documents the current contract.

Changes

• docs/concepts/reward-functions.rst: rewritten around reward_fn(record) -> float. Explains the prompt-RL record fields, the agentic-RL trajectory fields (messages, trace, tool_calls, tool_results, rendered_completion, final_answer), and how dataset-loader rows reach rewards via source_record / solutions -> answer. Points at the existing example reward files.
• docs/reference/reward-function-api.rst: documents the full RewardRecord field table, the scalar return contract, and the --reward-fn-path loading/validation behavior (must define callable reward_fn(record)).
• docs/troubleshooting/reward-function.rst: drops the batch-shape check ("reward list length matches the completions list length") in favor of the per-record contract.
• README.md, AGENTS.md, CONTRIBUTING.md: replace remaining reward_fn(example, completions) -> list[float] references. The README SDK quickstart additionally called the example reward file with the old shape, which no longer matches examples/math/math_verify_reward.py; it now builds one RewardRecord per rollout sample, mirroring what the CLI trainer does. These three files are outside docs/ but still presented the old shape as the current contract, so they are included to satisfy the acceptance criterion.

Verification

• grep across docs, README, AGENTS.md, CONTRIBUTING.md, CODEMAP.md, and examples: no reward_fn(example, completions) references remain.
• sphinx-build -b html -E docs builds with zero warnings and no broken internal links.
• The rewritten README quickstart snippet passes Python syntax compilation.
• Docs-only change; no code or public API changes.