Benchmark durations were originally formatted with higher precision
because they contained just the NsPerOp field. Now that we're
approximating the total duration they are more accurate and don't need
as many decimals.
The NsPerOp was previously casted to a time.Duration (int64) before
multiplying it with the number of iterations. This has been fixed by
doing the casting after the multiplication. The resulting duration
should be a bit more accurate.
