NM has a worker thread for every sliver ever to exist on the node. Each thread works off of a queue where the database adds a state from the statemachine to execute. To keep from spawning upwards of 300 worker processes during start(), a cummulative global delay is incremented (by 2) for every sliver known to the db. After the delay, start() is run. But, it's possible for the delay to grow without bound when NM is resynced to PLC and the db adds previously added states to the queue and causing the cumm start delay to be incremented once again. This patch keeps the queue from growing beyond the 4 known states. Any other additions are likely repeates and don't need to be added again. This is a bandaid; really we should keep a table of what exists in the queue and keep from readding while also bounding the delay to some maximum over the number of slivers present on the node.