for ref #3657.  The Oracle dialect makes more use of the "select_wraps_for"
feature than SQL server because Oracle doesn't have "TOP" for a limit-only
select, so tests are showing more happening here.  In the case where
the select() has some dupe columns, these are deduped from the .c collection
so a positional match between the wrapper and original can't use .inner_columns,
because these collections wont match.  Using _columns_plus_names
instead which is the deduped collection that determines the SELECT display,
which definitely have to match up.

(cherry picked from commit aa9ce3f521f254da9879ede011e520ec35b8270e)

would be turned into a list of individual characters.  This would
impact among other things using the :meth:`.Query.get` method
on a primary key that's a bytes object.
fixes #3660

the new behavior of the autoincrement flag as per ref #3216

handled within visit_select(); this attribute was added in the
1.0 series to accommodate the subquery wrapping behavior of
SQL Server and Oracle while also working with positional
column targeting and no longer relying upon "key fallback"
in order to target columns in such a statement.  The IBM DB2
third-party dialect also has this use case, but its implementation
is using regular expressions to rewrite the textual SELECT only
and does not make use of a "wrapped" select at this time.
The logic no longer attempts to reconcile proxy set collections as
this was not deterministic, and instead assumes that the select()
and the wrapper select() match their columns postionally,
at least for the column positions they have in common,
so it is now very simple and safe.  fixes #3657.
- as a side effect of #3657 it was also revealed that the
strategy of calling upon a ResultProxy._getter was not
correctly calling into NoSuchColumnError when an expected
column was not present, and instead returned None up to
loading.instances() to produce NoneType failures; added
a raiseerr argument to _getter() which is called when we
aren't expecting None, fixes #3658.

exactly

to not be loaded, if the joined eager load were from a row where the
same entity were present multiple times, some calling for the attribute
to be eagerly loaded and others not.  The logic here is revised to
take in the attribute even though a different loader path has
handled the parent entity already. fixes #3431

for example an exception object made within a test suite can
still repr (error seen in Keystone)