fill_record() has some logic to trigger the update of generated
columns for the tables to update. However, when the list of columns
to update is empty, no table will be identified for insert, and hence
the generated columns are never updated before sending the row to
the storage engine. No wonder all sorts of inconsistencies occur.

Looking more closely at calls to fill_record(), we see that we always
insert into exactly one table, and that table is always known by the
caller, hence it is a far better choice to add a TABLE * argument to
fill_record() and use it inside, instead of identifying the table
from the field arguments.

Doing this also triggered removing quite a bit of code.

A heads up to the update of TABLE::auto_increment_field_not_null:
It must be set to false only when the list of columns supplied to
the function is non-empty. I did not try to understand this logic,
just copied the existing logic as closely as possible.