Adaptive spam handling

[bclindner]

Being able to watch the public timeline is much more helpful if we have some way of inferring what is considered spam to an instance based on human input.

The actual filtering should be trivial - #11 satisfies that requirement - but this would require very manual input by the Ivory operator to train, which would be tedious. Ideally the spam handler should be able to take user input. The surrounding system would have to have a lot of moving parts, unfortunately - taking arbitrary user input and running with it is always problematic - but it's the only way we can truly count on an anti-spam measure moving forward.