Adding a new kind of relationship loader that is
a cross between the "immediateload" and the "subquery"
eager loader, using an IN criteria to load related items
in bulk immediately after the lead query result is loaded.

Change-Id: If13713fba9b465865aef8fd50b5b6b977fe3ef7d
Fixes: #3944

The ``lazy="select"`` loader strategy now makes used of the
:class:`.BakedQuery` query caching system in all cases.  This
removes most overhead of generating a :class:`.Query` object and
running it into a :func:`.select` and then string SQL statement from
the process of lazy-loading related collections and objects.  The
"baked" lazy loader has also been improved such that it can now
cache in most cases where query load options are used.

Change-Id: Ic96792fffaa045ae9aa0a4657d6d29235d3efb85
Fixes: #3954

psycopg2 has updated their unicode / string handling, greatly
reducing callcounts on Python 2.x.

Change-Id: Idffd242b303ffea9c7733068a4be290da4dbe679

Continuing from Ie43beecf37945b2bb7fff0aaa597a597293daa18,
also observed is the overhead of PathRegsitry memoized token
functions, as these paths are not cached in the case of
long joinedloader chains.  The memoizations here were made
with short paths in mind, and have been replaced with
an inlined straight create of these paths up front, producing
callcounts very similar to 0.8. Combined with the previous
optimizations, 1.1 now runs the "joined eager load of one row"
worst case test in about 40% fewer calls than 0.8 and 60%
fewer than 1.1.5.

Change-Id: Ib5e1c1345a1dd8edfbdb3fed06eb717d4e164d31
Fixes: #3915

Addressed some long unattended performance concerns within the
joined eager loader query construction system.  The use of ad-hoc
:class:`.AliasedClass` objects per query, which produces lots of column
lookup overhead each time, has been replaced with a cached approach
that makes use of a small pool of :class:`.AliasedClass` objects
that are reused between invocations of joined eager loading.
Callcount reduction of SQL query generation for worst-case joined
loader scenarios (lots of joins, lots of columns) is reduced by
approximately 270%.

Change-Id: Ie43beecf37945b2bb7fff0aaa597a597293daa18
Fixes: #3915

as subqueries in order to work around SQLite's lack of support for this
syntax, is lifted when SQLite version 3.7.16 or greater is detected.
fixes #3634
- The workaround for SQLite's unexpected delivery of column names as
``tablename.columnname`` for some kinds of queries is now disabled
when SQLite version 3.10.0 or greater is detected.
fixes #3633

profiling problems here
- add extras_require to setup.py for the most common DBAPIs
- rework tox.ini to use extras, specify a test matrix built in

method, and its interaction with result-row processing, now allows
the columns passed to the method to be positionally matched with the
result columns in the statement, rather than matching on name alone.
The advantage to this includes that when linking a textual SQL statement
to an ORM or Core table model, no system of labeling or de-duping of
common column names needs to occur, which also means there's no need
to worry about how label names match to ORM columns and so-forth.  In
addition, the :class:`.ResultProxy` has been further enhanced to
map column and string keys to a row with greater precision in some
cases.  fixes #3501
- reorganize the initialization of ResultMetaData for readability
and complexity; use the name "cursor_description", define the
task of "merging" cursor_description with compiled column information
as its own function, and also define "name extraction" as a separate task.
- fully change the name we use in the "ambiguous column" error to be the
actual name that was ambiguous, modify the C ext also