Don't count requeued test failures toward circuit breaker

[ianks]

Summary

the circuit breaker (max_consecutive_failures) was counting requeued test failures as real failures. this meant a worker that hit a few flaky tests in a row would get killed even though the tests were being retried, and if that was the last active worker, those requeued tests just sat in the queue forever. build would finish reporting 0 failures.

before (max_consecutive_failures=3):

                        breaker sees     reality
  test A fails          failure #1       requeued (handled)
  test B fails          failure #2       requeued (handled)
  test C fails          failure #3       requeued (handled)
                        KILL worker

  worker dead. tests are back in queue but no workers left.
  build finishes: 0 failures.

after:

                        breaker sees     reality
  test A fails          (nothing)        requeued (handled)
  test B fails          (nothing)        requeued (handled)
  test C fails          (nothing)        requeued (handled)
  ...
  test A fails again    failure #1       requeue exhausted
  test B fails again    failure #2       requeue exhausted
  test C fails again    failure #3       requeue exhausted
                        KILL worker

  worker dead, but 3 failures are actually recorded.

Changes

• moved report_failure! / report_success! after the requeue check in handle_test_result
• requeued tests are now invisible to the circuit breaker. they dont increment or reset the consecutive failure counter
• only tests that actually fail (requeue exhausted or not attempted) count

Tradeoff

The tradeoff here is that workers which are legitimately in a corrupted state will requeue more tests than before, making a test suite more likely to fail rather than a worker being killed. Ultimately, the tradeoff is that this "less resilient to worker corruption" behavior is better than the current "can silently skip failing tests" behavior.

[dpetrick]

Reasoning sounds good to me, I'd prefer someone with internal knowledge to confirm if possible, but I think it's good to go.