Open checkpoint might grow unbounded and lead to high allocation in key-index at replica

Repetition of the the same test as in MB-49134, this time on a toy-build that contains stats improvements by MB-48587.

Run: http://perf.jenkins.couchbase.com/job/oceanus/7373/
Build: 7.1.0-1670 + MB-48587

The test is a simple bulk load (via restore) of 320 Million items on a 2-KVNode cluster. 1 Bucket, 1 Replica.
We have a success from the test PoV, as all the items are successfully stored and replicated.
The problem is the checkpoints_overhead_estimate_replica stat (yellow line) that shows a constantly increasing asymptotic trend.
Charts show that the checkpoints_overhead_estimate_index_replica (dark-blue line) is what determines that trend.

Note
The above ran with the default checkpoint_memory_ratio=30%.
Interestingly enough, lowering the CMQuota (eg, 10%, http://perf.jenkins.couchbase.com/job/oceanus/7372/parameters) shrinks all the other mem-usage stats too, so the pattern stays the same.