Fix caching error in the push evaluator

Initialising `result` to `{}` in the parameters meant that every call to
_flatten_dict used the *same* target dictionary.

I'm hopeful this will fix https://github.com/matrix-org/synapse/issues/2270,
but I suspect it won't. (This code seems to have been here since forever,
unlike the bug, and I don't really think it explains the observed
behaviour). Still, it makes it hard to investigate the problem.