[codex] Add indexed Lance column fetch for interleaved batches

[VibhuJawa]

What changed

• add LanceDatasetConfig, LanceIndexCacheConfig, and the public LanceColumnFetchStage
• support presence-only exact-key lookup and arbitrary typed Lance column projections
• preserve input ordering while handling duplicate input keys, missing keys, and explicit destination-column policies
• keep a persistent Lance session, B-tree cache, and _take_rows adapter in an actor stage
• add InterleavedLanceReader and allow LanceWriter to accept Arrow-backed InterleavedBatch tasks
• document the API and add focused reader/writer/fetch tests

Why

Interleaved workflows need one reusable keyed-fetch primitive for image presence, binary payloads, metadata, hashes, and future annotation columns. Keeping the dataset identity separate from the projected payload removes per-row Lance locators and lets downstream workflows fetch only the columns they need.

Validation

• 24 passed across the new stage, inherited Lance reader/writer tests, and the unchanged Ray adapter regression suite
• 310 passed, 6 skipped in the broader interleaved suite; four unrelated existing log-capture assertions fail because Loguru output is not surfaced through caplog
• real version-pinned S3 smoke test against the stable-row-ID MINT image table and its 16-segment URL B-tree
• focused Ruff checks passed for all touched files

This PR intentionally excludes the MINT converter, Slurm workflows, pilot data-building scripts, and benchmark results.