RFR: Soft refs should be purged reliably on allocation failure, or with compact heuristics

[Aleksey Shipilev]

http://cr.openjdk.java.net/~shade/shenandoah/softrefs-trigger/webrev.01/

Have been following up on the question from Simon: why Full GC cleans up several times more heap
than degenerated or normal cycle? My gut feeling was about soft-references, and experiments
confirmed it. But, this led me to discover our soft ref policy needs reliability fixes. This patch
does them: we purge soft refs on allocation failure, and some heuristics (compact) do that
unconditionally.

Testing: IDEA with sh/jdk8 and this patch beckported, tier3_gc_shenandoah (running)

Thanks,
-Aleksey