These flake occasionally because registering a request increases
the current request count synchronously while unregistering
decreases it asynchronously, meaning if the test has time to
call register 100 times before unregister can run it will
wrongfully flake.
This could be addressed in production code but as comments in
the change explain, this isn't worth the complexity. Hence
just have a pinpoint fix in the test. See aosp/2707373 for
what a fix in the production code would look like.
Test: manual
Bug: 289530922
Change-Id: Iad9a725eda91406f820abe4706bca0a4756352a4
2eb2f7998d