Fix the issue where the first measurement in a microbenchmark is

slightly slower, which is especially noticeable with quick benchmarks.

See the measurements (column A) from the
TrivialKotlinBenchmark#nothing microbenchmark from 100 runs before and
after this change in
https://docs.google.com/spreadsheets/d/1W6oxK_rM0KGzZBLxpax05766-hYyV2EYLw49dIs2uPo/edit?usp=sharing&resourcekey=0-3rhUgg9eSBekHXF8mOznng.

At head, every first measurement of a run (rows 1, 51, 101, 151, etc)
are in the range of ~5.5-6ns, while the median is 3.6ns. With this
change (the benchmark-first-measurement tab), this issue disappears.

The standard deviation at head is 0.42, while with this change it
reduces to 0.19.

Test: Existing tests.
Change-Id: Id6f1dac88d19665e9fa80d8e60144f38089f4cc2

benchmark/benchmark-common/src/main/java/androidx/benchmark/BenchmarkState.kt

@@ -319,6 +319,11 @@ class BenchmarkState internal constructor(phaseConfig: MicrobenchmarkPhase.Confi
             profilerResult = phaseProfilerResult
         }
 
+        // Warm up the metrics data structure to reduce the impact on the first measurement.
+        currentMetrics.captureStart()
+        currentMetrics.captureStop()
+        currentMetrics.captureInit()
+
         currentMetrics.captureStart()
         return true
     }