feat: batch generic row lookups across base and WAL

[beinan]

Adds GenericStore::get_many, the shared Rust trait/facade/client API, and POST /api/v1/generic/{name}/get-rows. Read up to 1,024 IDs together or project only id for deduplication; missing IDs are omitted and repeated IDs returned once in first-requested order.

Lance 9 serializes disk lookups for bare PK-IN queries. A documented zero-offset compatibility adapter selects its general indexed LSM scan, retaining flushed-WAL/newest-key semantics. Reads reuse merge-consistency retries and avoid source-local limits that could underfill results. A known Lance 9 WAL membership-index error retries native point lookups against the same captured view after discarding partial results. That compatibility path is slower; other errors still propagate. No dependencies are added or upgraded. Remove unused lance-graph and lancedb dependencies from core; the former pulled a second Lance/DataFusion stack, legacy TLS/JWT dependencies, and Thrift into consumers of the batch API.

Validation: 22 core and 10 HTTP route tests pass, including a live Rust-client round trip, a single-index-plan regression, WAL overwrites, base deletions, projections, and concurrent merges. Clippy passes for all affected crates/targets.

Local ARM64 debug benchmarks, five warm samples, 512 IDs:

• Private 721-row conversation fixture: full-row P50 3,703.85 → 42.35 ms; ID-only 18.87 ms.
• Synthetic base + two pending WAL generations: full-row P50 7,892.15 → 49.42 ms.

Every response matched expected rows. These are local warm-cache query measurements, not Blob or ingestion throughput. The PR includes a reproducible synthetic benchmark and methodology/results in docs/benchmarks/batch-get.md; private data is not included.