RFR: 8241081: Shenandoah: Do not modify update-watermark concurrently

[Roman Kennke]

JDK-8240873 introduced short-cutting arraycopy-barriers on objects
beyond the update-watermark. Concurrently updating the update-watermark
after a region has been completely updated proves problematic: we see
various different crashes related to this. Strengthening the fences
around it doesn't fully solve the problem, apparently the problem is
deeper. It doesn't seem worth to keep doing this for very little gain.

Bug:
https://bugs.openjdk.java.net/browse/JDK-8241081
Webrev:
http://cr.openjdk.java.net/~rkennke/JDK-8241081/webrev.02/

Testing: hotspot_gc_shenandoah and the failing TestStringDedupStress,
both in a loop for some hours

Thanks,
Roman