noderoster: optimize event list query

[benbroadaway]

Current query applies a possibly very big join against the process_queue table. That can lead to never catching up if the task falls behind for some reason. This change filters-fast down to the configured limit before joining the process_queue data.

Some future considerations:

• Add a task run limit (e.g. run duration or number of events processed) so a slow server has a termination point to allow a roll of the dice for another server to
  • currently, a "slow" server may run the task indefinitely
• Tweak the host cache. Use weight instead of duration for eviction. Or maybe just lengthen the time beyond a minute
  • When bogged down, a single loop through events may take more than a minute (or two).
  • Either way, make it configurable so OOM-prone combo can be changed without code change