Increase in tmp_ooms due to checkpoint memory causes drop in throughput

We're seeing a ~9% regression in HiDD->S0->Update only tput test (98k ops/s to 89k ops/s) - when running without perfrunner backups (i.e. load.use_backup.false passed as an override parameter to the Jenkins job).

I've bisected down the regression, looks to be due to this commit (went in 7.2.0-1515):

https://github.com/couchbase/kv_engine/commit/52bb81c5a470c4d682952909209fb4cbac87668a

I tried reverting that commit and applying it over a recent build (7.5.0-3265) and the regression goes away:

• baseline build 7.5.0-3265 (89k ops/s): http://perf.jenkins.couchbase.com/job/rhea-5node1/3391/

                                                                                                      VS

• toy manifest of 3265 + reverting that commit (98k ops/s): http://perf.jenkins.couchbase.com/job/rhea-5node1/3392/

Here's the bisection journey and the comparison between 7.2.0-1514 and 7.2.0-1515 where the increase in tmp_ooms is first seen and checkpoint_mem is seen to reach the 30% ratio: https://issues.couchbase.com/browse/MB-54242?focusedCommentId=637086&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-637086