Shenandoah still does concurrent cleanup in control thread w/o parallelization, it usually takes few ms in a cycle to process all the trashed regions with single thread.

Meanwhile current impl takes heap-lock when it recycle trashed regions, which makes it even slower and causes contention on heap-lock.

After making concurrent cleanup parallelized, we may also reset marking bitmap, so in next cycle we only need to reset the rest active regions.