DOC: Add contributor instructions for dataset loaders

[romanlutz]

Description

Adds .github/instructions/datasets.instructions.md with the conventions Copilot (and human contributors) should follow when adding or modifying seed dataset loaders under pyrit/datasets/seed_datasets/.

Covers the points that recur in dataset PR reviews:

• SeedObjective vs SeedPrompt — picking the right seed type for behavior/goal rows vs literal messages
• Subclassing _RemoteDatasetLoader and using its helpers (_fetch_from_huggingface, _fetch_from_url, _validate_enum / _validate_enums)
• HuggingFace gating: token argument + HUGGINGFACE_TOKEN env fallback, docstring requirements
• Enum-based filters (no raw strings / Literals), validated via the inherited helpers
• Per-seed metadata preservation and class-level metadata constants picked up by _parse_metadata
• Raising ValueError on empty filter results
• Registration in remote/__init__.py (imports + __all__)
• Bibliography wiring (doc/references.bib + the hidden-citations block in doc/bibliography.md)
• Test conventions (mock _fetch_from_huggingface, asyncio-auto, token forwarding)
• Optional live sanity-check against the real HF dataset before opening the PR

Modeled on the structure, depth, and frontmatter style of converters.instructions.md, scenarios.instructions.md, and output.instructions.md. Does not duplicate rules already in style-guide.instructions.md (async suffix, kw-only, type hints, enums-over-Literals).

Tests and Documentation

Markdown-only contributor-instructions change — no code, tests, or notebooks affected.