perf: delete builds the dataset-wide row id index only to translate ids it does not need

Description

On a dataset with stable row ids, DeleteJob scans _rowid, captures the ids of the rows its predicate matched, and then builds the dataset-wide RowIdIndex through get_row_id_index purely to translate those ids back into the row addresses the deletion vectors need. Deleted rows keep no ids, so the round trip is avoidable: the scan can capture _rowaddr directly. On a dataset without stable row ids _rowid already is the address, so the same addresses come out; only the accumulator changes.

Measured on a local (not memory://) dataset with --profile release-no-lto: building the index cold costs 0.6 ms at 500 fragments / 50k rows and 3.6 ms at 5000 fragments / 100k rows. The whole one-row delete on the 500-fragment dataset takes about 47 ms, so that cost is not visible. The win is a simplification whose size grows with fragment count; it is not a significant speedup. Where nothing else in the plan needs it, capturing _rowaddr also keeps a delete of a few rows from holding a dataset-wide structure in memory. On object storage the same change would avoid reading fragments' external row-id metadata; the local fixture keeps that metadata inline, so the measurement does not exercise that path.

Steps to reproduce

// A delete whose predicate does not fold to a literal takes this path:
//   DeleteJob::execute_impl -> scanner.with_row_id().project(&[ROW_ID])
//   -> make_rowid_capture_stream -> get_row_id_index(&dataset)
//   -> CapturedRowIds::row_addrs(index)
let mut dataset = /* dataset written with enable_stable_row_ids: true */;
dataset.delete("id = 12345").await.unwrap();

Expected behavior

DeleteJob resolves the addresses it needs without building the row id index itself.

Lance version

13.0.0-beta.3 (main)

Language binding

Rust

Environment

macOS ARM, local storage.

Start at DeleteJob::execute_impl and follow scanner.with_row_id().project(&[ROW_ID]) into make_rowid_capture_stream, get_row_id_index, and CapturedRowIds::row_addrs. Check how the capture can use _rowaddr directly for stable and non-stable row ids. Done means DeleteJob resolves deletion addresses without building the dataset-wide RowIdIndex and the shown delete path still works.