Materialize deletions when updating scalar indices

When we update a scalar index we basically retrain it on the old data (which we scan from the old index, already sorted) and the new data (which we scan from the dataset, in sorted order). While we are scanning the old data we scan both the value column and the row id column. We can grab the deletion information for the affected fragments from the dataset and filter out rows based on the row id. This should allow us to relatively cheaply materialize deletions at the same time we run an update.

Start at the scalar index update path and trace how old index data and new dataset data are scanned. Check how deletion information for affected fragments can be used with row IDs; done means deletions are materialized during the update and deleted rows are filtered from the retraining input.