FEAT Add Turkish Prompt-Injection dataset loader - #2365

Adds a remote SeedDatasetProvider loader for the open, CC BY 4.0 Turkish Prompt-Injection dataset (3nesdeniz/turkish-prompt-injection-1k), pinned to an immutable HF revision.

Turkish-native prompt-injection coverage is scarce in the existing seed datasets. This loader adds paired benign/attack Turkish examples (instruction override & system-prompt extraction, jailbreak personas, obfuscation & data exfiltration, and agentic tool abuse) with Turkish morphological variation. The benign hard negatives share surface vocabulary with the attacks while being fully legitimate, so false-positive behavior can be measured directly.

Changes:

• _TurkishPromptInjectionDataset with typed label (attack/benign/all), split, and attack_families filtering, pinned to an immutable HF revision, with schema and label-dependent validation. Follows the established _RemoteDatasetLoader pattern.
• Registers the loader and its enums for discovery.
• Adds 11 focused unit tests (mocked HF fetch + validation/error cases, including the type(label) is int bool-rejection guard).
• Adds a bibliography/reference entry.

Testing:

• pytest of the new test file: 11 passed.
• Live-loaded the published test split end-to-end (79 attacks / 34 benign) and verified split and family filters against real data.
• black, isort, and flake8 clean.

Labels describe prompt-injection decisions rather than content harms, so they are retained in seed metadata and harm_categories is intentionally empty (matching the agentic boundary-pairs loader convention).