mean vs welfordMean

[harendra-kumar]

Add tests to demonstrate the difference between mean and welfordMean. We can use alternate small and large numbers with a high dynamic range for testing the loss of precision. A Double has 52 significand bits and 10 exponent bits. We can design a test based on that knowledge. See https://en.wikipedia.org/wiki/Floating-point_arithmetic .

It would be nice to have sum operations that works on a wider range possibly by sacrificing some computational complexity e.g. multiple buckets for different precision levels.

[harendra-kumar]

As per the current benchmarks welfordMean is significantly faster than mean:

      mean:                     OK (0.20s)
        3.08 ms ± 173 μs
      welfordMean:              OK (0.22s)
        859  μs ±  44 μs

[harendra-kumar]

The perf situation has changed after recent changes, mean is faster now.