Improve handling of uncommitted DB txns with "uncaught" exceptions
One of the causes (if not the cause) of bug 56269 is if something
manages to throw an exception that makes it to MediaWiki::run's
last-resort exception catcher while having an open database transaction:
the transaction never gets committed or rolled back, so it gets
implicitly rolled back and a warning is raised.
The API has the opposite problem in bug 63145: it catches the exception
but then does the normal DB shutdown which *commits* the transaction.
This is certainly the Wrong Thing to do.
Ideally, neither of these would ever happen because any code using
transactions would have its own try-catch that would catch any
exception, rollback the transaction, and then rethrow the exception. But
it happens anyway, so let's have both of these last-resort exception
handlers do the rollback, and also log the exception so the throwing
code has a better chance of being properly fixed.