[STF] Adapt timing_with_fences test to be more reliable (#2658)

* Change the timing_with_fences test to be compatible with a system that is very busy

* Do not make expectations about timers

Author: caugonnet

cudax/test/stf/utility/timing_with_fences.cu

@@ -65,13 +65,19 @@ void run(int NTASKS, int ms)
 
   ctx.finalize();
 
-  float elapsed;
+  [[maybe_unused]] float elapsed;
   cuda_safe_call(cudaEventElapsedTime(&elapsed, start, stop));
 
-  // Give 25% margin of error
-  float expected = 1.0f * NTASKS * ms;
-  float err      = fabs(elapsed - expected) / expected;
-  EXPECT(err < 0.25f);
+  [[maybe_unused]] float expected = 1.0f * NTASKS * ms;
+
+  /* We cannot really expect this measurement to be accurate because the
+   * thread(s) executing the code might be preempted on a system with a high load
+   * (as during unit tests). So the best we can expect is that the elapsed time
+   * is larger than the sleep time, but event the timer on the GPU is not
+   * perfectly accurate so we do not make any strict assumptions about the
+   * test, and just keep this test to demonstrate how to use the mechanisms,
+   * and ensure they are functional . */
+  // EXPECT(elapsed >= expected);
 }
 
 int main(int argc, char** argv)