Add score_answer versioning and historical scorer resolution

[sileod]

Motivation

• Provide stable, auditable scorer behavior by recording scorer version, content hash, and commit with each generated example to allow reproducible scoring.
• Allow loading legacy score_answer implementations by version, hash, or from a historical commit/file so that scorer changes remain backward-compatible.
• Surface the change in the task authoring guide so task authors know to bump versions when changing scorers.

Description

• Add a new reasoning_core/score_answer_history.py module to compute scorer hashes, locate repository commits, load historical scorer source from files or git commits, resolve callables, and return the appropriate scoring function via resolve_score_answer_fn.
• Extend reasoning_core/template.py to declare Task.score_answer_version and Task.score_answer_history, provide helpers score_answer_hash, resolve_score_answer_fn, and score_answer_for_entry, and record _score_answer metadata (version, hash, commit) in generate_example().
• Update reasoning_core/__init__.py to invoke the new per-entry scoring entrypoint via DATASETS[task_name].score_answer_for_entry(...) so scorers chosen per-entry are applied correctly.
• Update TASK_AUTHORING_GUIDE.md to document the _score_answer metadata and the recommended workflow for bumping scorer compatibility.
• Add unit tests in tests/test_score_answer_versioning.py covering default metadata recording, loading a legacy scorer from a file via score_answer_history, and error behavior when a requested legacy version is not registered.

Testing

• Executed pytest tests/test_score_answer_versioning.py which ran the new tests and they passed.
• The tests validate that _score_answer metadata is recorded, legacy scorer files can be loaded for historical version, and that missing score_answer_history entries raise a KeyError as expected.

---

Codex Task