Choose between horizontal and vertical merge insert based on data width

Currently there are two approaches we can take during merge insert.

A horizontal merge insert does not create a new fragment. It tombstones the columns that are being modified and rewrites those columns. This introduces write amplification based on the columns being updated (the entire column(s) must be rewritten).

A vertical merge insert marks the updated rows deleted and writes new rows. This introduces write amplification based on the columns not being updated (since they must be rewritten into the new rows).

Currently we choose horizontal merge insert if the incoming data is a partial schema and vertical merge insert if the incoming data is the full schema. However, this is not always the correct choice. For example, a user created a table where an image column started out as null and then started to backfill the image column. Since they were only filling in the image column this led to a tremendous amount of write amplification because a horizontal merge insert was being used.

We could potentially choose the correct strategy based on the sizes of the columns being modified.

Start by tracing the existing horizontal and vertical merge insert paths and how they determine the incoming schema and column sizes. Define and validate a strategy-selection rule that reduces write amplification, with tests showing the chosen approach for partial updates such as backfilling one column and for full-schema updates.