pep-249 exception names linked to exception classes of an entirely
different name, preventing SQLAlchemy's own exception wrapping from
wrapping the error appropriately.
The SQLAlchemy dialect in use needs to implement a new
accessor :attr:`.DefaultDialect.dbapi_exception_translation_map`
to support this feature; this is implemented now for the py-postgresql
dialect.
fixes #3421

functionality.   Added a new "soft invalidate" feature to the
connection pool at the level of the checked out connection wrapper
as well as the :class:`._ConnectionRecord`.  This works similarly
to a modern pool invalidation in that connections aren't actively
closed, but are recycled only on next checkout; this is essentially
a per-connection version of that feature.  A new event
:class:`.PoolEvents.soft_invalidate` is added to complement it.
fixes #3379

- Added new flag
:attr:`.ExceptionContext.invalidate_pool_on_disconnect`.
Allows an error handler within :meth:`.ConnectionEvents.handle_error`
to maintain a "disconnect" condition, but to handle calling invalidate
on individual connections in a specific manner within the event.

- Added new event :class:`.DialectEvents.do_connect`, which allows
interception / replacement of when the :meth:`.Dialect.connect`
hook is called to create a DBAPI connection.  Also added
dialect plugin hooks :meth:`.Dialect.get_dialect_cls` and
:meth:`.Dialect.engine_created` which allow external plugins to
add events to existing dialects using entry points.
fixes #3355

fail to store the correct value for MSSQL on an INSERT where the
primary key value was present in the insert params before execution.
fixes #3360

That is, after exhausing all rows using the fetch methods, the
DBAPI cursor is released as before and the object may be safely
discarded, but the fetch methods may continue to be called for which
they will return an end-of-result object (None for fetchone, empty list
for fetchmany and fetchall).   Only if :meth:`.ResultProxy.close`
is called explicitly will these methods raise the "result is closed"
error.
fixes #3330 fixes #3329

persistence.py could theoretically hit the limit of the cache
(100 items by default) and at some points fail to have a key that
we check for, due to the cleanup.   This has never been observed
so its likely that so far, the total number of INSERT, UPDATE and
DELETE statement structures in real apps has not exceeded 100
on a per-mapper basis; this could happen for apps that run a
very wide variety of attribute modified combinations into the unit
of work, *and* which have very high concurrency going on.
This change will be a lot more significant when we open up
use of LRUCache + compiled cache with the baked query extension.

The "wrapping" employed by the mssql and oracle dialects using the
"iswrapper" argument was not being used intelligently by the compiler,
and the result map was being written incorrectly, using
*more* columns in the result map than were actually returned by
the statement, due to "row number" columns that are inside the
subquery.   The compiler now writes out result map on the
"top level" select in all cases
fully, and for the mssql/oracle wrapping case extracts out
the "proxied" columns in a second step, which only includes
those columns that are proxied outwards to the top level.

This change might have implications for 3rd party dialects that
might be imitating oracle's approach.   They can safely continue
to use the "iswrapper" kw which is now ignored, but they may
need to also add the _select_wraps argument as well.

such that they are matched to the received result set positionally,
rather than by name.  Originally, this was seen as a way to handle
cases where we had columns returned with difficult-to-predict names,
though in modern use that issue has been overcome by anonymous
labeling.   In this version, the approach basically reduces function
call count per-result by a few dozen calls, or more for larger
sets of result columns.  The approach still degrades into a modern
version of the old approach if textual elements modify the result
map, or if any discrepancy in size exists between
the compiled set of columns versus what was received, so there's no
issue for partially or fully textual compilation scenarios where these
lists might not line up.  fixes #918
- callcounts still need to be adjusted down for this so zoomark
tests won't pass at the moment

:meth:`.Connection.invalidate` method, or an invalidation due
to a database disconnect, would fail if the
``isolation_level`` parameter had been used with
:meth:`.Connection.execution_options`; the "finalizer" that resets
the isolation level would be called on the no longer opened connection.
fixes #3302