One bounded slice of a dump, plus where to resume.
Phases run ddl, rows, later, done; indexes come after the data (each is a charged row per insert).
ddl
rows
later
done
DUMP_START to begin, then whatever the previous chunk returned
if cursor.shape disagrees with opts (two different dumps would be spliced)
cursor.shape
opts
One bounded slice of a dump, plus where to resume.
Phases run
ddl,rows,later,done; indexes come after the data (each is a charged row per insert).