FEAT: Add remote WildGuardMix dataset loader

[romanlutz]

Summary

Adds a remote seed-dataset loader for AllenAI's WildGuardMix moderation corpus. The previously proposed TrustAIRLab remote loader has been removed from this PR; that corpus is being evaluated separately as a static, credited jailbreak-template collection.

The AdvBench investigation also confirmed that no new remote AdvBench loader is needed. Its separate harmful_strings.csv gap is tracked in #1826.

WildGuardMix loader

Loader	Source	Rows	Size	License	Access
_WildGuardMixDataset	allenai/wildguardmix	86,759 train + 1,725 test	huge	ODC-BY 1.0	Hugging Face token plus acceptance of the separate AI2 Responsible Use Guidelines

The loader:

• Fetches both wildguardtrain and wildguardtest configs by default, with a splits filter to narrow the result.
• Defaults to harmful, adversarial jailbreak prompts. Vanilla prompts, benign companions, and train rows containing responses remain explicit opt-ins through prompt_harm_labels, adversarial, and prompt_only.
• Maps all 14 harmful upstream subcategory slugs to canonical PyRIT HarmCategory names while preserving the raw source label in seed metadata. Explicitly unharmful rows remain uncategorized.
• Is tagged {safety, jailbreak} and intentionally not tagged default.
• Accepts an explicit token or falls back to HUGGINGFACE_TOKEN.

AdvBench investigation

PyRIT's local adv_bench.prompt contains 520 rows with harm-category metadata and matches the canonical llm-attacks/data/advbench/harmful_behaviors.csv goals. The viable Hugging Face mirror contains only 500 rows, so replacing or duplicating the local dataset with a remote loader would regress coverage.

Canonical AdvBench also includes 574 output-side strings in harmful_strings.csv, which serve a different scorer-evaluation use case. That gap is tracked separately in #1826.

Validation

• 36 focused WildGuardMix unit tests cover both configs, filtering, empty results, enum validation, metadata, all upstream harm-category mappings, and token forwarding.
• All 1,181 dataset unit tests pass after removing TrustAIRLab from this branch.
• Daily dataset e2e discovery includes WildGuardMix and skips cleanly when no token is configured or the account has not accepted the dataset terms.
• The complete set of 14 harmful source subcategories was checked against a public WildGuardMix mirror; unit-test prompts are fabricated mocks rather than copied gated rows.

Documentation and licensing

• Adds the han2024wildguard citation and includes WildGuard in the built-in dataset documentation and rendered provider list.
• WildGuardMix data is licensed under ODC-BY 1.0. The AI2 Responsible Use Guidelines are a separate Hugging Face access gate.
• PyRIT does not redistribute the corpus; the loader fetches it at runtime after the user obtains access.

Follow-ups

• Add AdvBench harmful_strings local dataset (574 output strings) #1826 — add a local dataset for AdvBench harmful_strings.csv.
• TrustAIRLab in-the-wild prompts — assess each deduplicated prompt separately for conversion into a reusable {{ prompt }} jailbreak template.

[adrian-gavrila]

Approved with one comment on tests, feel free to resolve if you disagree there and merge