Fix formatting durations with fractions with leading zeroes

[bens]

The duration formatting function gives incorrect results for values with leading zeroes in the fraction component, for example I see this result:

benchmark              runs     total time     time/run (avg ± σ)     (min ... max)                p75        p99        p995      
-----------------------------------------------------------------------------------------------------------------------------
1 sphere               10       20.979s        2.97s ± 80.398ms       (2.7s ... 2.294s)            2.133s     2.70s      2.70s      

The total time is 20.9 seconds, but when divided by ten for the average it gives 2.9 seconds, not 2.09. Also see the min, max, and percentiles not making sense. This does change the output for cases like 1.0s to 1.000s, but I think it's better to show the precision in that case anyway.

Another issue was the division was always floored since it was using integer division, I've changed to using a float for the division down to a nice unit and using format's builtin rounding off at the time of printing.

Finally, for nanoseconds there's no actual fractional precision, so I added a special case for that to only show whole nanoseconds.

I've added tests for these cases as well of course.

[bens]

Output from the same benchmark after fixing the code makes a lot more sense:

benchmark              runs     total time     time/run (avg ± σ)     (min ... max)                p75        p99        p995      
-----------------------------------------------------------------------------------------------------------------------------
1 sphere               10       20.553s        2.055s ± 37.139ms      (2.015s ... 2.134s)          2.091s     2.064s     2.064s     

[bens]

Alternatively, just using the stdlib's duration formatting is a good option too IMHO:

pub fn duration(buffer: []u8, ns: u64) ![]u8 {
    return try std.fmt.bufPrint(buffer, "{}", .{
        std.fmt.fmtDuration(ns),
    });
}

[hendriknielaender]

Good point. Leveraging the std lib can reduce complexity and potential errors in this custom implementation.

[hendriknielaender]

@bens 🌟 Great work on your first contribution, and welcome to the project!

Thanks for adjusting the tests and fixing this.

[bens]

No idea why, but this PR seems to have become disconnected from my branch and isn't reflecting changes there. Will make a new one.