fix(index): keep the trailing rows when shuffling a v2 buffer

[jackylee-ch]

IvfShuffler::total_batches divides a v2 unsorted buffer's row count by SHUFFLE_BATCH_SIZE, so the trailing partial batch is never read; the legacy arm uses the exact reader.num_batches(). A buffer shorter than one batch counts as zero batches and write_partitioned_shuffles returns no files at all, while the caller goes on to commit an index whose fragment bitmap claims full coverage. shuffle_vectors is the only constructor passing is_legacy = false, reached from IndicesBuilder.shuffle_transformed_vectors, so a distributed index build silently loses those vectors.

Round the batch count up, and clamp the row range both readers derive from it -- ReadBatchParams::Range rejects a range past the end of the file.

Testing

cargo test --release -p lance-index --all-features shuffler::test. The new test_v2_shuffle_keeps_every_row covers a partial trailing batch, a buffer shorter than one batch, and an exact multiple; on the old arithmetic the first loses 7 of 2055 rows and the second produces no output. write_unsorted_stream only emits v1 files, so the fixture writes the v2 buffer directly.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The v2 shuffler now schedules the trailing logical batch and clamps both read passes to the file’s true row count, preserving every transformed vector without changing the legacy path. The regression covers a partial tail, a sub-batch file, and an exact multiple.