fix: chunked index stream to work around arrow's 2Gb limits

[kushudai]

Chunk BTree index Arrow handoff to stay under Arrow's 2 GiB offset cap

What

• The range-BTree build buffers a whole fragment group into one narrow VectorSchemaRoot. A value column exceeding 2 GiB overflows Arrow's 32-bit-offset VarChar/VarBinary buffer during accumulation.
• Add ArrowBatchChunker: accumulates rows into multiple narrow batches, rolling before any var-width column's data buffer crosses the 2 GiB cap (plus a row-count ceiling so fixed-width/offset buffers stay under it). Fails fast on a single value over the cap.
• Add ChunkedArrowReader: SingleBatchArrowReader generalized to emit N pre-built batches; caller owns/closes them.
• Rewire buildForFragmentGroup to chunked accumulation. lance-core re-zones the multi-batch stream, so the built index is identical to a single-batch handoff; peak memory is unchanged.

Notes

• Both new classes are schema/index-agnostic. Only the range-BTree path is wired as it's the only build that hands a Spark-side Arrow stream to lance-core (withPreprocessedData); FTS/zonemap/fragment-BTree read from the dataset directly and never buffer a VectorSchemaRoot.

Test plan

• Unit (JUnit 5): ArrowBatchChunkerSuite (byte-budget split, single batch, over-budget isolation, fixed-width row cap, order/content) and ChunkedArrowReaderSuite (ordered emission, exhaustion).
• Integration (Spark 3.5, native Lance): AddIndexTest range-mode — create, repeat/replace, rows_per_range, index-details — all pass.
• spotless:apply + checkstyle:check clean.

[github-actions]

ACTION NEEDED
Lance follows the Conventional Commits specification for release automation.

The PR title and description are used as the merge commit message. Please update your PR title and description to match the specification.

For details on the error please inspect the "PR Title Check" action.