feat(lance): push LIMIT down into Lance scans

[jja725]

Summary

Adds LIMIT pushdown to the Presto Lance connector. PrestoDB has no applyLimit SPI, so this uses a ConnectorPlanOptimizer.

• Plan optimizer: LanceLimitPushdown rewrites a LimitNode sitting directly over a Lance TableScanNode into a LanceTableHandle carrying the row count (wired via LanceConnectorPlanOptimizerProvider → LanceModule → LanceConnector.getConnectorPlanOptimizerProvider). Only fires when LIMIT is directly over the scan (no intervening FILTER), so semantics stay exact.
• Split coalescing: with a pushed-down limit, the split manager coalesces just enough leading fragments to cover it into a single split (deletion-aware via Fragment.metadata().getNumRows()) instead of one-split-per-fragment.
• Scan limit: the fragment page source passes ScanOptions.limit(n) to stop each scan early.

Best-effort IO reduction — the engine still enforces the exact row count.

Test plan

• New unit tests: TestLanceLimitPushdown (pushes into scan, idempotent), TestLanceSplitManager (split-per-fragment vs single coalesced split + coalesce helper edge cases), additions to TestLanceTableHandle (limit field + JSON round-trip).
• ./mvnw -o test -pl presto-lance → 37 tests, BUILD SUCCESS.

Stacked on #4 — review/merge that first. The diff here is just the LIMIT-pushdown changes.

🤖 Generated with Claude Code