feat(rewards): support batched reward-hook execution (#225)

[candy972023]

• ✨ New feature
• ⚡ Performance improvement

Summary

Closes #225

Add an optional batch reward interface alongside the existing per-example adapter. Users can define reward_fn_batch(records) -> list[float] in the same module as reward_fn; when present, AReno calls the batch interface once per batch instead of looping per record. This benefits reward functions with heavy per-call overhead (e.g. external model scoring, batch inference).

Files

File	Type	Description
areno/api/rewards.py	Modified	Add load_reward_fns, validate_batch_rewards, compute_rewards with cardinality validation and execution timing
areno/api/trainers/policy_only.py	Modified	Use compute_rewards in both agentic and rollout paths
areno/api/trainers/ppo.py	Modified	Forward reward_fn_batch to PolicyOnlyTrainer
areno/api/trainer_factory.py	Modified	Accept and forward reward_fn_batch
areno/cli/train.py	Modified	Load both functions via load_reward_fns
tests/test_batched_reward_cpu.py	New	15 CPU tests

Design

• Dual interface: reward_fn(record) -> float (existing) and reward_fn_batch(records) -> list[float] (optional). Users opt in by defining the batch function; no config change needed.
• Cardinality validation: validate_batch_rewards checks output length matches input length. On mismatch, raises ValueError with the batch index so the caller can identify which prompt-group batch failed.
• Execution timing: Both paths record elapsed time via time.perf_counter() and log it through the trainer's logger.
• Backward compatibility: load_reward_fn preserved as public API, delegates to load_reward_fns(path)[0]. When no batch function is defined, behavior is identical to before.
• Architecture: Reuses existing RewardRecord contract, no new dependencies.

How was it tested?

CPU tests (no GPU required), 15/15 passed:

• Batch and per-example paths produce identical results on a deterministic hook
• Cardinality validation: short output, long output, batch index in error message
• Empty batch returns empty list
• Float coercion of integer rewards
• load_reward_fns loads both functions, detects absent batch fn, rejects non-callable
• Backward compatibility: dispatch without batch fn matches manual per-example loop
• Execution timing logged for both paths

Checklist

• Default behavior remains backward compatible
• No external database or mandatory sandbox
• Uses existing AReno contracts (RewardRecord, Callable)
• No new dependencies
• Focused CPU tests cover success, invalid input, and boundary cases