fix: Rare race in cluster tests

[godexsoft]

While working on #3174 we had a failure in https://github.com/XRPLF/clio/actions/runs/33074493704/job/98527800974?pr=3174 which appears to be an actual race. This PR fixes the small and rare race that the run exposed.

Backend::run() starts two independent RepeatedTasks on separate strands: a reader
(doRead → onNewState_) and a writer (doWrite → writeNodeMessage). Seven tests
asserted EXPECT_CALL(writeNodeMessage).Times(AtLeast(1)) but only ever waited on the
read path (semaphore.acquire() is released from the state callback). When the test body
returns, ~Backend → stop() sets state_ = Stopped, and RepeatedTask's loop checks
state_ after the timer wait — so a writer whose first tick hasn't landed yet breaks
out without ever calling doWrite(). CI hit exactly that: writeNodeMessage
"never called - unsatisfied and active".

Both intervals are 1 ms, so the writer almost always wins the race — hence the rarity.

• Relaxed those seven expectations to Times(AnyNumber()). None of them inspects the
  written message; the expectation only existed because the fixture is a StrictMock.
  The write path stays properly covered by the two WriteNodeMessage* tests, which
  release the semaphore from inside the write action and so actually wait for it.
  Stop keeps AtLeast(1) (it sleeps 20 ms and asserting both tasks ran is its point).
• std::binary_semaphore → std::counting_semaphore<> in the fixture: both callbacks
  are WillRepeatedly on a 1 ms timer, so a second release() before the main thread's
  acquire() was UB.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!