XDCR + Rebalance-in test shows variation in Sherlock vs. 3.1.0

This bug used to be MB-16086.

Title of the test on showfast:

"Rebalance-in (min), 4->5 > 5>5 UniDir, 1 bucket x 1G x 1KB, 10K ops/sec"

Repeat runs show that variations only exist in Sherlock and not in 3.1.0.

(order of the values here follows showfast order, not actual test execution order):

With 3.1.0:
113.4
111.1
116.6
(one more data point in progress 8/24)

With Sherlock RC4 4047:
177.6
225.9
69.9
164.8

The last two runs were obtained such that the versions are alternated (run1 with 3.1.0, run2 with RC4, run3 with 3.1.0, etc ... ).