Make speed_benchmark_torch report latency in us (#37953)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37953

Earlier it said us but reported ms.

Test Plan: buck run aibench:run_bench -- -b aibench/specifications/models/pytorch/fbnet/fbnet_mobile_inference.json --devices s9u --remote --framework pytorch --logger_level info --job_queue aibench_interactive --platform android/full_jit

Reviewed By: xcheng16

Differential Revision: D21349612

fbshipit-source-id: b97b6216eb0264123ff2c7852a0678b2008b0bf1
diff --git a/binaries/speed_benchmark_torch.cc b/binaries/speed_benchmark_torch.cc
index 13069ea..b3c4126 100644
--- a/binaries/speed_benchmark_torch.cc
+++ b/binaries/speed_benchmark_torch.cc
@@ -211,23 +211,23 @@
       ".");
   caffe2::Timer timer;
   std::vector<float> times;
-  auto millis = timer.MilliSeconds();
+  auto micros = timer.MicroSeconds();
   for (int i = 0; i < FLAGS_iter; ++i) {
     auto start = high_resolution_clock::now();
     module.forward(inputs);
     auto stop = high_resolution_clock::now();
-    auto duration = duration_cast<milliseconds>(stop - start);
+    auto duration = duration_cast<microseconds>(stop - start);
     times.push_back(duration.count());
   }
-  millis = timer.MilliSeconds();
+  micros = timer.MicroSeconds();
   if (FLAGS_report_pep) {
     for (auto t : times) {
       std::cout << "PyTorchObserver {\"type\": \"NET\", \"unit\": \"us\", \"metric\": \"latency\", \"value\": \"" << t << "\"}" << std::endl;
     }
   }
-  std::cout << "Main run finished. Milliseconds per iter: "
-            << millis / FLAGS_iter
-            << ". Iters per second: " << 1000.0 * FLAGS_iter / millis
+  std::cout << "Main run finished. Microseconds per iter: "
+            << micros / FLAGS_iter
+            << ". Iters per second: " << 1000.0 * 1000 * FLAGS_iter / micros
             << std::endl;
 
   return 0;