Use multiple threads to process ParallelGC deferred updates

The "Deferred updates" step during full GCs takes a significant fraction of the total compaction time. It should be possible to parallelise this, although care needs to be taken not to introduce regressions if there are no/few large objects to update.

See here: https://mail.openjdk.org/pipermail/hotspot-gc-dev/2022-August/039905.html