The current implementation of checking for validating the number of rows
for computed columns for in non-IMPORT INTO backups is inefficient. This
commit moves the check from being performed on every row to only being
performed once per import.

Release note: None

This commit moves the computation of finding the expected number of data
columns to the creation of the input converter, rather than
re-calculating it on every row. The memory footprint of loading all of
the visible columns for every row was noticeable.

Release note (performance improvement): CSV imports should now be
slightly faster.