fix(data): warn when behavioral dataset config keys are silently ignored

[tianyi-zhang-02]

What does this PR do ?

load_response_dataset instantiates with dataset_class(**data_config), and every built-in loader accepts **kwargs — so a standard ResponseDatasetConfig key the class doesn't support is silently swallowed. E.g. subset: socratic on gsm8k silently loads the hardcoded main config, and split_validation_size/seed silently do nothing on the 12 loaders that never call split_train_validation (while the SFT guide documents them as generic keys).

This adds a dispatcher-level guard, warn_on_unsupported_dataset_config_keys, that warns when a user-set behavioral key (subset, split, split_validation_size, seed) is not declared anywhere in the resolved class's __init__ MRO.

Design notes:

• MRO-aware: kwargs-forwarding subclasses that delegate to a declaring base (the intent datasets) are not flagged.
• functools.partial registry entries (AIME variants) are unwrapped before inspection.
• None values and falsy split_validation_size are skipped, so documented config boilerplate (subset: null, split_validation_size: 0) stays quiet — none of the shipped examples/configs emit warnings.
• Warning, not error, for backward compatibility.
• Known limitation: an external dotted-path dataset class that consumes one of these keys via kwargs.get(...) without declaring it in any __init__ signature would see a spurious warning. None of the built-in loaders do this (verified), and it's warning-only by design — but flagging it for review. Scoped to the response dispatcher; the preference dispatcher has the same pattern and can adopt the shared helper as a follow-up.

Issues

Closes #3270. Generalizes the failure mode @RayenTian flagged in the #3131 review.

Usage

data:
  train:
    dataset_name: gsm8k
    subset: socratic   # before: silently loads "main"; after: UserWarning at dispatch

Before your PR is "Ready for review"

• New tests: 7 unit cases for the guard (unsupported key warns; supported key quiet; None/0 quiet; MRO forwarding quiet; partial unwrapped; non-class input safe) plus an end-to-end dispatcher test, in tests/unit/data/datasets/test_response_dataset_registry.py (registry-dispatch scope).
• Validation: helper cases verified locally (all pass); ruff check / ruff format --check clean; pyrefly==0.24.2 (CI-pinned) reports no new errors on the two whitelisted files touched.
• Docs: N/A (warning makes existing documented behavior observable; no doc changes required).

cc @RayenTian @yuki-97

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.