When running a benchmark multiple times, instead of re-computing the
value of b.N each time, use the value found by the first run.

For

  go test -bench=. -benchtime 3s -count 2 p_test.go

on the benchmark in the linked issue; before:

  BenchmarkBenchmark-4   	     500	  10180593 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.111079ms
  	  p_test.go:13: single call took 1.017298685s
  	  p_test.go:13: single call took 5.090096124s
  BenchmarkBenchmark-4   	     500	  10182164 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.098169ms
  	  p_test.go:13: single call took 1.017712905s
  	  p_test.go:13: single call took 5.090898517s
  PASS
  ok  	command-line-arguments	12.244s

and after:

  BenchmarkBenchmark-4   	     500	  10177076 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.091301ms
  	  p_test.go:13: single call took 1.016943125s
  	  p_test.go:13: single call took 5.088376028s
  BenchmarkBenchmark-4   	     500	  10171497 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.140245ms
  	  p_test.go:13: single call took 5.085605921s
  PASS
  ok  	command-line-arguments	11.218s

Fixes #23423

Change-Id: Ie66a8c5ac43881eb8741e14105db28745b4d56d3
Reviewed-on: https://go-review.googlesource.com/110775Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>