feat(eval): retrieval-quality eval harness for code_index

[itomek]

Context

Reviewer on #721 asked for testing evidence on large repos and retrieval quality. The PR ships unit tests covering parser correctness and SDK behaviour, but there is no end-to-end retrieval-quality eval.

Goal

A gaia eval code-index subcommand that runs a curated query set against a reference repository (e.g. gaia itself), measures recall@k / MRR, and outputs a report consumable by gaia report.

Acceptance

• New eval subcommand wired into src/gaia/eval/ alongside existing fix-code / agent evals.
• Curated query set with known-correct answers committed under tests/fixtures/.
• Baseline numbers committed to docs/ so regressions are visible.
• Report renders via gaia report.

References

• src/gaia/eval/ — existing eval framework
• docs/reference/eval.mdx — framework docs
• src/gaia/code_index/sdk.py — SDK under test

Deferred from #721.