Use time.perf_counter_ns instead of time.time_ns to have sub-microsecond precision on both Linux and Windows (#3220)

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

python/triton/testing.py

@@ -54,7 +54,7 @@ def __init__(self, **kwargs):
         self.record()
 
     def record(self):
-        self.timestamp = time.time_ns() / 1_000_000
+        self.timestamp = time.perf_counter_ns() / 1_000_000
 
     def elapsed_time(self, end):
         return end.timestamp - self.timestamp