Maintain fractional millisecond precision when recording wall time ev…

…ents (#1658)

Use fp division instead of integer division to maintain sub-ms
precision. Without this patch, the tutorials fail w/ upstream pytorch on
small shapes because the runtime is less than 1ms and we get a divide by
zero error in processing the timing.

cc #1563

python/triton/testing.py

@@ -22,7 +22,7 @@ def __init__(self, **kwargs):
             self.record()
 
         def record(self):
-            self.timestamp = time.time_ns() // 1_000_000
+            self.timestamp = time.time_ns() / 1_000_000
 
         def elapsed_time(self, end):
             return end.timestamp - self.timestamp