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.