This new technique is more straightforward and compatible with test
parallelization, where the effective database connection settings no
longer match settings.DATABASES.

Rewrote the test for #9479 according to the original ticket.

It contained models, forms and admin objects causing undesirable
import side effects. Refs #16368.

Thanks to Ramiro, Carl and Loïc for the review.

Signed-off-by: Jason Myers <jason@jasonamyers.com>

Should be unneeded with Python 2.7 and up.
Added some unicode_literals along the way.

Fixed #20483.

enter_transaction_management() was nearly always followed by managed().

In three places it wasn't, but they will all be refactored eventually.
The "forced" keyword argument avoids introducing behavior changes until
then.

This is mostly backwards-compatible, except, of course, for managed
itself. There's a minor difference in _enter_transaction_management:
the top self.transaction_state now contains the new 'managed' state
rather than the previous one. Django doesn't access
self.transaction_state in _enter_transaction_management.

There were a couple of errors in ._dirty flag handling:
  * It started as None, but was never reset to None.
  * The _dirty flag was sometimes used to indicate if the connection
    was inside transaction management, but this was not done
    consistently. This also meant the flag had three separate values.
  * The None value had a special meaning, causing for example inability
    to commit() on new connection unless enter/leave tx management was
    done.
  * The _dirty was tracking "connection in transaction" state, but only
    in managed transactions.
  * Some tests never reset the transaction state of the used connection.
  * And some additional less important changes.

This commit has some potential for regressions, but as the above list
shows, the current situation isn't perfect either.