Bucket compaction causes massive slowness of flusher and UPR consumers

5 -> 5 UniDir, 2 buckets x 500M x 1KB, 10K SETs/sec, LAN

Similar to MB-11731 which is getting worse and worse. But now compaction affects intra-cluster replication and XDCR latency as well:

"ep_upr_replica_items_remaining" reaches 1M during compaction
"xdcr latency" reaches 5 minutes during compaction.

See attached charts for details. Full reports:

http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=atlas_c1_300-1005_a66_access
http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=atlas_c2_300-1005_6d2_access

One important change that we made recently - http://review.couchbase.org/#/c/39647/.

The last known working builds is 3.0.0-988.