Support key-only delete for primary-key tables

[qusijun]

Problem

For primary-key tables, delete records currently go through the normal row-based write path as RowKind::Delete.

In practice, this seems to require callers to provide a full row, even when the delete is logically identified only by the primary key.

Why this is an issue

For common delete-by-PK use cases, this can be inefficient:

• callers may only know the primary key, not the full previous row
• if non-PK fields are required, callers may need to read the old row first
• this adds unnecessary read, serialization, and write-path overhead

This makes the current behavior closer to a changelog-style row delete than a true key-only delete.

Expected behavior

It would be useful to support key-only delete for PK tables, so callers can delete using only primary key fields without reconstructing the full row.

Possible directions:

• add a dedicated key-only delete API
• allow delete input with only PK fields
• represent such deletes internally as tombstones / key-only delete records
• optimize the path to avoid full-row materialization where possible