feat: add retrieval-quality + version A/B evaluation harness (#98)

[dcfocus]

Summary

Adds a retrieval-quality evaluation harness that measures search/retrieve quality against a labeled query set and compares quality across dataset versions — closing #98. Cross-version regression eval is the differentiator a stateless vector DB can't offer.

This PR is independent of #108 (#100) and #109 (#102) — it branches from main and can be reviewed/merged on its own.

Core (lance-context-core::eval)

• Eval dataset format — EvalQuerySet / EvalQuery / RelevanceLabel, referencing records by stable external_id with optional graded relevance, loadable from / serializable to JSONL.
• Metrics — recall@k, precision@k, MRR, nDCG@k (linear gain), hit-rate, as pure functions.
• Runner — ContextStore::evaluate(query_set, config) scores a set at the current version for vector (search) or hybrid (retrieve) mode, with configurable k, filters, and lifecycle options. Returns per-query + aggregate scores and a manifest (query-set id, version, k, mode, distance metric).
• Version A/B — ContextStore::evaluate_versions(..., baseline, candidate) checks out each version, evaluates, restores the original version, and reports per-metric deltas. MetricScores::delta also enables config A/B (two evaluate runs).

Python

• Context.evaluate(queries, *, k, mode, filters, include_expired, include_retired) → report dict.
• Context.evaluate_versions(queries, baseline_version, candidate_version, ...) → {baseline, candidate, deltas}.

Query sets and reports are marshaled as JSON across the pyo3 boundary, reusing the serde-derived core types.

Tests

• core (12): metric correctness vs hand-computed fixtures (perfect / rank-2 / none / graded nDCG / precision@k), JSONL round-trip, vector + hybrid runners, filter and lifecycle handling, and the version-A/B mechanism (zero-delta + version restore).
• python (5): test_eval.py — vector mode, graded nDCG, config A/B delta, version A/B, invalid-mode error.

README gains an eval + version-A/B example.

Checks

• cargo test -p lance-context-core --lib — 59 passed
• cargo fmt --all -- --check and cargo clippy --workspace --all-targets -- -D warnings — clean
• ruff format --check, ruff check, pyright — clean
• test_eval.py — 5 passed

Acceptance criteria (#98)

• Run a labeled query set → recall@k / MRR / nDCG for vector and hybrid retrieval
• A/B two versions (or two configs) → per-metric deltas
• Results pinned to a context version + config in a reproducible report
• Tests cover metric correctness against fixtures, filter/lifecycle handling, and the version-A/B mechanism

Notes

• Cross-version A/B with differing data is not unit-tested because MemWAL-backed writes don't reliably advance the base-table manifest version (the same limitation test_time_travel_checkout xfails on). The A/B mechanism (checkout both versions, restore, per-metric deltas) and metric-difference detection are covered via the same-version and config-A/B tests.
• The eval-set format is intentionally simple JSONL so it can be shared with the post-training pipeline's eval/holdout set (Curate lance-context records into trainable datasets (SFT/preference/rollout export) #96 / Reproducible train/eval split helper for post-training exports #103), per the issue notes.
• nDCG uses linear gain (grade / log2(rank+1)); documented on MetricScores.

Closes #98