ORM attributes can now be assigned any object that is has a
``__clause_element__()`` attribute, which will result in inline
SQL the way any :class:`.ClauseElement` class does.  This covers other
mapped attributes not otherwise transformed by further expression
constructs.

As part of this, it was considered that we could add
__clause_element__() to ClauseElement, however this causes endless loops
in a "while" pattern and this pattern has been identified in third
party libraries.  Add a test to ensure we never make that change.

Change-Id: I9e15b3f1c4883fd3909acbf7dc81d034c6e3ce1d
Fixes: #3802

An adjustment to ON CONFLICT such that the "inserted_primary_key"
logic is able to accommodate the case where there's no INSERT or
UPDATE and there's no net change.  The value comes out as None
in this case, rather than failing on an exception.

Change-Id: I0794e95c3ca262cb1ab2387167d96b8984225fce
Fixes: #3813

The "raise_on_sql" option differentiates from "raise" in that
firing a lazy loader is OK as long as it does a simple
get from identity map.   Whereas "raise" is more useful
for the case that objects are to be detached.

As part of this, refactors the strategy initiation logic
a bit so that a LoaderStrategy itself knows what "key" was used
to create it, thus allowing variants of a single strategy
based on what the "lazy" argument is.  To achieve this we
have to also get rid of _get_strategy_by_cls().

Everything here is internal with the one exception of an apparently
undocumented, but not underscored, "strategy_class" key
on relationship().   Though it's not clear what
"strategy_class" accomplishes; at this point the strategy
system is extensible using Property.strategy_for().

Fixes: #3812
Change-Id: I812ad878ea5cf764e15f6f71cb39eee78a645d88

The change in #3431 still checks that the instance() is
non-None, deferring to other loading schemes if it is.
These columns are dedicated towards the entity however, so if the value
is None, we should set it.  If it conflicts, we are detecting that
in any case.

Change-Id: I223768e2898e843f953e910da1f9564b137d95e4
Fixes: #3811

Changed the query used to get "default schema name", from one that
queries the database principals table to using the
"schema_name()" function, as issues have been reported that the
former system was unavailable on the Azure Data Warehouse edition.
It is hoped that this will finally work across all SQL Server
versions and authentication styles.

Change-Id: Ic11bd4162c0d6a60432ae44876e86512703c1f81
Fixes: #3810

Updated the server version info scheme for pyodbc to use SQL Server
SERVERPROPERTY(), rather than relying upon pyodbc.SQL_DBMS_VER, which
continues to be unreliable particularly with FreeTDS.

Change-Id: I4ff49ae13c8ff51bd764980131d41c18d73d87ce
Fixes: #3814

Add checks in spots where state.obj() might be late-GC'ed before
we get a chance to call the event.   There may be more cases
of these which we should address as they come up.  The Session
should always be maintaining strong refs to objects that have
pending operations left on them, so for these cases we need
to ensure that ref remains long enough for the event to be called.

Change-Id: I1a7c7bc57130acc11f54ad55924af2e36ac75101
Fixes: #3808

Added compiler-level flags used by Postgresql to place additional
parenthesis than would normally be generated by precedence rules
around operations involving JSON, HSTORE indexing operators as well as
within their operands since it has been observed that Postgresql's
precedence rules for at least the HSTORE indexing operator is not
consistent between 9.4 and 9.5.

Fixes: #3806
Change-Id: I5899677b330595264543b055abd54f3c76bfabf2