JDK-8321815 introduced code to sync up thread states at safepoint. This code runs with the time proportional to the number of threads, and can take tens/hundreds of microseconds.

Unfortunately, that sync code runs right in `VMOperation::doit`, which is not captured in GC event time (done by `ShenandoahPausePhase` in `ShenandoahHeap::entry_*`). It is also not captured captured by "GC net time" in regular Shenandoah counters.

 It would be better to move these down in relevant `ShenandoahHeap::op_*` methods, and wrap them in GC phase counters to clearly capture the time we spend there.