Transaction loop should use explicit transaction manager mode

[jamadden]

This triggers better polling behaviour in a ZODB Connection (such as used with https://github.com/NextThought/nti.site/blob/master/src/nti/site/runner.py#L141)

[jamadden]

Of course, if we do it in our __call__, the connection that was opened by our subclass in nti.site doesn't get to take full advantage of that (because it opens the connection before calling us). Sigh.

So we need some extra hooks here to make that possible and/or we need to modify nti.site (e.g., to move opening the connection down into run_handler instead of __call__ — but having the same connection be used in all retry attempts is probably the right behaviour as it'll have the best cache).