Optimize ExecutionStateSampler

Showed up as 1.5% of CPU on Nexmark 11 streaming benchmark run.

It appears to be using a ConcurrentSkipListSet and most of the cpu is relatedd to inserts/removes in that, involving the system hash of the entries as that is used for ordering.

The consistent ordering is unnecessary. Additionally for other reasons, removal and iteration is already synchronized and so performance will likely be better just using a synchronized HashMap and synchronizing in the add case.