perf: narrow scan projection in merge_insert partial schema to avoid heavy column I/O

[hfutatzhanghb]

Summary

When performing merge_insert with a partial schema (e.g. only key + embedding columns for updates), the table scan in the join phase exposes ALL dataset columns. While DataFusion's optimizer can push projections down, the target."column_name" synthetic column references in the post-join loop effectively block this optimization — the planner sees references to every dataset column and must scan them all.

For large tables (1B rows, 20TB) where only 50K embedding values need updating, this means the join scans 20TB of data unnecessarily.

Fix

When a partial schema is detected (source_field_names.len() < self.dataset.schema().fields.len()), narrow the target scan to only _rowid, _rowaddr, and on-columns before the join via DataFrame::select_columns. DataFusion pushes this projection down to the table scan, avoiding I/O on heavy data columns during matching.

The target."column_name" synthetic column references in the post-join loop are unaffected — DataFusion's projection-pushdown operates at the physical-plan level and does not invalidate logical column references.

Performance Impact

	Before	After
Scan columns during join	All (20TB for 1B-row table)	_rowid + _rowaddr + on-columns (~16GB)
Reduction	—	~1000x less I/O for the join phase

Testing

• cargo check -p lance --lib passes
• cargo fmt --all passes