FEAT: Add CoCoNot refusal-calibration dataset loaders

[romanlutz]

Adds the CoCoNot dataset (Brahman et al. 2024, allenai/coconot, ODC-BY 1.0) as two sibling loaders, closing a long-standing gap in PyRIT's refusal-calibration coverage.

CoCoNot evaluates the inverse of typical refusal scoring: "did the model refuse a benign prompt it shouldn't have refused?" — catching over-aligned models. Five categories: incomplete, unsupported, indeterminate, humanizing, safety.

Loaders

Class	dataset_name	HF config / splits	Rows	Semantic
_CoCoNotRefusalDataset	coconot_refusal	original.train + original.test	12,478	Model should refuse.
_CoCoNotContrastDataset	coconot_contrast	contrast.test	379	Benign — model should comply.

• Refusal sibling accepts a splits: list[CoCoNotSplit] filter (multi-split iteration follows msts_dataset.py:194-201).
• Both siblings accept categories: list[CoCoNotCategory].
• AI2's reference noncompliant responses (populated only in original.train) preserved via metadata["response"].
• Contrast sibling intentionally not tagged "safety" since its prompts are benign.
• pref config (DPO pairs) deferred — targets training, not red-teaming.

Out of scope

Dedicated scorer, RefusalCalibration scenario, and cross-cutting "calibration" tag — the existing TrueFalseInverterScorer(SelfAskRefusalScorer(...)) already covers "compliance = success" semantics.

Validation

• 17 unit tests + full 511-test datasets suite green.
• Live HF check: counts match (12,478 / 379); category + split filters verified end-to-end (e.g. SAFETY → 3,531, TRAIN → 11,477, HUMANIZING + TRAIN → 1,795).
• Pre-commit hooks (ruff-format, ruff-check, ty-check) all pass.
• Bibliography entry brahman2024coconot added; doc/code/datasets/1_loading_datasets.{py,ipynb} regenerated.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com