* Add a bunch of profiling, so we can figure out where performance bottlenecks are.
* Cache the in-object token cache in memcached, bringing per-condition run times from 4ms to a cooler 296us
* Disallow more than 100 condition-checks per action.
* Disallow filters which are misbehaving (more than 5 matches, constituting more than 5% of recent actions)
* Display statistics for all filters (percentage of times the condition limit is reached, and percentage of times some filter is matched), and individual filters (percentage of times it's matched).