The docstrings for connect() and on_connect() were incorrect
between Dialect vs. DefaultDialect.   Redocumented related
methods, clean up formatting, and remove unicode-related
attribute descriptions from the top level Dialect document
as these don't apply to Python 3.

Change-Id: I45baab757f8e20627eba42c30b9e8dbe26356275
(cherry picked from commit 649de79950dcf952d7a44069faf36925c23c4e63)

Change-Id: I3ef36bfd0cb0ba62b3123c8cf92370a43156cf8f
(cherry picked from commit 699272e4dcb9aa71ebbc0d9487fb6de82d3abc2b)

Changed the mechanics of :class:`.ResultProxy` to unconditionally
delay the "autoclose" step until the :class:`.Connection` is done
with the object; in the case where Postgresql ON CONFLICT with
RETURNING returns no rows, autoclose was occurring in this previously
non-existent use case, causing the usual autocommit behavior that
occurs unconditionally upon INSERT/UPDATE/DELETE to fail.

Change-Id: I235a25daf4381b31f523331f810ea04450349722
Fixes: #3955
(cherry picked from commit 8ee363e4917b0dcd64a83b6d26e465c9e61e0ea5)

Change-Id: I4e8c2aa8fe817bb2af8707410fa0201f938781de

The "extend_existing" option of :class:`.Table` reflection would
cause indexes and constraints to be doubled up in the case that the parameter
were used with :meth:`.MetaData.reflect` (as the automap extension does)
due to tables being reflected both within the foreign key path as well
as directly.  A new de-duplicating set is passed through within the
:meth:`.MetaData.reflect` sequence to prevent double reflection in this
way.

Change-Id: Ibf6650c1e76a44ccbe15765fd79df2fa53d6bac7
Fixes: #3861

This allows to skip buffering of the results on the client side, e.g.
the following snippet:

    table = sa.Table(
        'testtbl', sa.MetaData(),
        sa.Column('id', sa.Integer, primary_key=True),
        sa.Column('a', sa.Integer),
        sa.Column('b', sa.String(512))
    )
    table.create(eng, checkfirst=True)

    with eng.connect() as conn:
        result = conn.execute(table.select().limit(1)).fetchone()
        if result is None:
            for _ in range(1000):
                conn.execute(
                    table.insert(),
                    [{'a': random.randint(1, 100000),
                      'b': ''.join(random.choice(string.ascii_letters) for _ in range(100))}
                      for _ in range(1000)]
                )

    with eng.connect() as conn:
        for row in conn.execution_options(stream_results=True).execute(table.select()):
            pass

now uses ~23 MB of memory instead of ~327 MB on CPython 3.5.2 and
PyMySQL 0.7.9.

psycopg2 implementation and execution options (stream_results,
server_side_cursors) are reused.

Change-Id: I4dc23ce3094f027bdff51b896b050361991c62e2

Compiler can now set up execution options and additionally
will propagate autocommit from embedded CTEs.

Change-Id: I19db7b8fe4d84549ea95342e8d2040189fed1bbe
Fixes: #3805

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

Tests illustrate that exceptions like GreenletExit and
even KeyboardInterrupt can corrupt the state of a DBAPI
connection like that of pymysql and mysqlclient.   Intercept
BaseException errors within the handle_error scheme and
invalidate just the connection alone in this case, but not
the whole pool.

The change is backwards-incompatible with a program that
currently intercepts ctrl-C within a database transaction
and wants to continue working on that transaction.  Ensure
the event hook can be used to reverse this behavior.

Change-Id: Ifaa013c13826d123eef34e32b7e79fff74f1b21b
Fixes: #3803