Reference cycle in _TracebackLogger and bug in _TracebackLogger.__del__() #155

Hi,

While trying to change how Trollius store tracebacks in Future.set_exception(), 
I found two bugs:

* _TracebackLogger object is never deleted if its activate() method is not 
called. It happens when run_until_complete() is used: the call to the 
activate() method is scheduled, but it will never be executed. It happens in 
many tests (test_tasks.py, test_futures.py, I don't remember which one)

* _TracebackLogger.__del__() does nothing if self.tb is None whereas self.exc 
is still known

The following patch fixes these two bugs:
http://codereview.appspot.com/69350045

Remarks:

- the patch modifies Future.__del__() to have the same output on Python 3.4: 
just add a trailing colon to the error message
- TestLoop.close() closes also BaseEventLoop.close() to clear _ready: it is 
needed in the new unit test to remove the last reference to the _TracebackLogger
- Future.set_exception()

By default (debug mode disabled), Future.set_exception() now clears the 
__traceback__ attribute of the exception to fix the reference cycle issue. So 
the log of unhandled exception doesn't contain the traceback anymore. It makes 
the debug much harder :-( But it avoids uncollectable objects in production.

To get the traceback, enable the debug mode of the event loop, but you might 
create uncollectable objects again.

I don't know what is the best choice: keep the traceback or avoid uncollectable 
objects. Is it common to use run_until_complete() and miss the call the 
activate()?

At least, the change of _TracebackLogger.__del__() must be applied.

--

Even with the patch, test_tasks.test_wait_errors() still creates uncollectable 
objects (Future, frames, TimerHandle, etc.) with Python 3.3. It's probably a 
different bug.