feat: add lightweight graph quality benchmarks

[Someday-czy]

Summary

• Add deeprefine benchmark prepare, evaluate, and report CLI commands.
• Add deterministic intrinsic and downstream graph-quality metrics.
• Add Re-DocRED-mini and 2Wiki-mini preparation adapters.
• Add a self-authored synthetic test fixture for offline metric and CLI verification.
• Add package data, attribution documentation, and benchmark tests.

Validation

• python -m pytest -q — 49 tests passed.
• python -m deeprefine_skill.cli benchmark --help
• python -m build

Scope

• Default evaluation uses the Python standard library and does not require a GPU or LLM API.
• The synthetic suite is only an expected fixture output for testing metric calculations and CLI behavior.
• No real Re-DocRED or 2Wiki performance result is claimed in this PR.
• Third-party raw datasets are not redistributed.
• G-BERTScore is optional and remains outside the default installation and CI path.

Notes

• evaluate reads saved baseline and candidate graphs without modifying them.
• Optional prediction JSONL files enable retrieval and answer metrics.