Support settings columns with expressions in merge_insert

Support queries like:

MERGE INTO <dataset>
USING <source>
ON dataset.id = source.id
WHEN MATCHED AND target.text <> source.text THEN UPDATE
  SET text = source.text,
      updated_at = timestamp_now(),
      text_embedding = embedding(source.text)
WHEN NOT MATCHED THEN INSERT
  (text, created_at, updated_at, text_embedding)
  VALUES
  (source.text, timestamp_now(), timestamp_now(), embedding(source.text))

This would allow deferring computations, such as embedding functions, until we find they are actually needed.

Start by locating the merge_insert implementation and tracing how its MATCHED UPDATE and NOT MATCHED INSERT clauses handle settings. Use the SQL example as the behavioral target, including expressions such as timestamp_now() and embedding(source.text), and verify that computation is deferred until the affected rows require it.