[Windows][Pytorch Upstream] Triton do_bench returns time not greater than 0 on Windows.

[etaf]

Describe the bug

The timing from triton's do_bench is not greater than 0 on windows.
The error:

(xinanlin) C:\xinanlin\pytorch>python test\inductor\test_benchmarking.py -vf
test_benchmark_cpu_smoke_benchmarker_cls0 (__main__.TestBenchmarker) ... inductor [('benchmarking.Benchmarker.benchmark_cpu', 1)]
ok
test_benchmark_cpu_smoke_benchmarker_cls1 (__main__.TestBenchmarker) ... inductor [('benchmarking.TritonBenchmarker.benchmark_cpu', 1)]
ok
test_benchmark_gpu_smoke_benchmarker_cls0 (__main__.TestBenchmarker) ... inductor [('benchmarking.Benchmarker.benchmark_gpu', 1)]
expected failure
test_benchmark_gpu_smoke_benchmarker_cls1 (__main__.TestBenchmarker) ... inductor [('benchmarking.TritonBenchmarker.benchmark_gpu', 1), ('benchmarking.TritonBenchmarker.triton_do_bench', 1)]
FAIL

======================================================================
FAIL: test_benchmark_gpu_smoke_benchmarker_cls1 (__main__.TestBenchmarker)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "c:\xinanlin\pytorch\torch\testing\_internal\common_utils.py", line 3108, in wrapper
    method(*args, **kwargs)
  File "c:\xinanlin\pytorch\torch\testing\_internal\common_utils.py", line 557, in instantiated_test
    test(self, **param_kwargs)
  File "c:\xinanlin\pytorch\torch\testing\_internal\common_utils.py", line 840, in test_wrapper
    return test(*args, **kwargs)
  File "C:\xinanlin\pytorch\test\inductor\test_benchmarking.py", line 82, in test_benchmark_gpu_smoke
    self.assertGreater(timing, 0)
AssertionError: -1.3166919946670532 not greater than 0

To execute this test, run the following from the base repo dir:
    python test\inductor\test_benchmarking.py TestBenchmarker.test_benchmark_gpu_smoke_benchmarker_cls1

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

----------------------------------------------------------------------
Ran 4 tests in 0.592s

FAILED (failures=1, expected failures=1)

The benchmark use Triton's do_bench as follow:
https://github.com/pytorch/pytorch/blob/6bdf2addc5e83cdf67af1261c417b3c91104ad02/torch/_inductor/runtime/benchmarking.py#L181-L202

Environment details

Triton main 6abfacb
PyTorch main 12662901aa8ced7d1118193e40dd4f018bb2e0fa