What does this PR do ?

Closes #2497.

Two related improvements to the validate() function in GRPO and Distillation, bundled into one PR because they touch the same function in the same way.

1. Per-task validation metrics (main change)

When data.validation is configured as a list of multiple datasets the multi-validation path correctly loads them all and dispatches per-task to the right environment during rollout, but the validation aggregator collapses everything into a single sample-weighted accuracy and avg_length. Per-task progress (e.g. gsm8k vs math500) is silently lost.

task_name is already on every sample (DatumSpec.task_name, preserved through rl_collate_fn into val_batch[\"task_name\"]); validate() simply did not read it. This PR teaches both validate() functions to track rewards per task during the loop and emit:

• accuracy_<task> for each task seen
• num_samples_<task> for each task seen

The aggregated accuracy key is preserved unchanged so existing dashboards continue to work. Single-task runs and legacy datasets without task_name are unaffected (the per-task block is skipped).

The driver-log summary also gains a per-task block:

📊 Validation Results:
    • Accuracy: 0.5957
    • Average response length: 512.8 tokens
    • Samples processed: 1819
    • Per-task accuracy:
        - data-math500: 0.4320 (n=500)
        - gsm8k: 0.6580 (n=1319)

2. max_val_samples becomes optional, Distillation truncation matches GRPO

Bundled because the patches sit a few lines apart in the same validate() block.

• grpo.GRPOConfig.max_val_samples was already typed as int | None (grpo.py:150) for NeMo-Gym compatibility but the main path crashed on None. Distillation's TypedDict required int outright. Both now accept None/absent and fall back to the full val_dataloader.
• Distillation's truncation switches from ceiling division to floor division, matching GRPO. Behaviour change is bounded to runs whose max_val_samples is not divisible by val_batch_size; all shipped recipes under examples/configs/recipes/llm/ use values that divide cleanly so none are affected.

Files touched

Out of scope

• DPO already emits per-dataset metrics via its dict[str, StatefulDataLoader] architecture (see nemo_rl/algorithms/dpo.py:332-377, prefix validation-{dataset_name}). No change needed.
• NeMo-Gym path in GRPO is left untouched. It clobbers val_batch_size to len(val_dataset) at setup time, which is specific to its single-batch eval and not something the main paths should adopt.
• Exemplar YAMLs (examples/configs/grpo_math_1B.yaml, examples/configs/distillation_math.yaml) keep their explicit max_val_samples so the recommended default stays documented.

Backwards compatibility

• Aggregated validation/accuracy key unchanged in value.
• Single-task validation: per_task_rewards ends up with one key, you get one extra validation/accuracy_<the-only-task> metric. Harmless.
• Datasets without task_name: per-task block is skipped, behaviour matches the old code.

Issues

Closes #2497.

Usage

A recipe with multiple validation datasets now reports per-task metrics:

data:
  validation:
    - dataset_name: gsm8k
      split: test
    - dataset_name: ResponseDataset
      data_path: data/math500.parquet
distillation:
  val_batch_size: 8
  # max_val_samples omitted -> evaluate the entire val dataset

wandb will plot validation/accuracy, validation/accuracy_gsm8k, and validation/accuracy_ResponseDataset separately. The existing aggregated validation/accuracy panel keeps working unchanged.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• Followed the config-conventions skill: optional max_val_samples field expressed via NotRequired, no hidden defaults, exemplar YAMLs keep their explicit recommended values.
• Two commits, one per logical change, to make the per-task and the max_val_samples patches reviewable independently.
• Open question for reviewers: should task-name slugification be added (e.g. data-math500 → data_math500)? Hyphens work in wandb but render slightly oddly in some downstream stores. Current patch keeps the name verbatim; happy to slugify if preferred.
• Branched off origin/main in a separate worktree to keep the change isolated from fix: pass trust_remote_code=True to remaining AutoConfig.from_pretrained sites #2496 (fix/autoconfig-trust-remote-code); the two PRs touch disjoint files.