Adding MCSE and ESS for more estimators

[sethaxen]

We could in principle support estimating the ESS for quite a few more estimators:

• var
• percentile
• Base.Fix1{mean}
• geomean
• harmmean
• genmean
• skewness
• kurtosis
• Base.Fix2{moment}
• mad

All but the first 3 require identifying suitable expectand proxies and overloading _expectand_proxy accordingly. In every case, the proxy needs to be rigorously tested with a simulation study on distributions with various shapes. I worked on this some time ago for several of these, so this is mostly a note for self to finish that work and add the results here.

We can similarly overload mcse using the same proxy, where we generally need to do some extra work to estimate the variance of the estimator. This would in general be faster than our mcse fallback and probably be less likely to underestimate the MCSE.

[sethaxen]

Here are the proposed expectand proxies for common scalar statistics in Statistics and StatsBase.

Need no simulation study:

• variation/sem: same as std
• Base.Fix2(percentile, p): same as Base.Fix2(quantile, p)
• Base.Fix1(mean, f): x -> f.(x)

Need a simulation study:

• var: same as std
• skewness: x -> (x .- mean(x)).^3
• kurtosis: x -> (x .- mean(x)).^4
• Base.Fix2(moment, p): x -> (x .- mean(x)).^p
• geomean: x -> log.(x)
• harmmean: x -> inv.(x)
• genmean: x -> iszero(p) ? log.(x) : x.^p
• mad: x -> (y=abs.(x .- median(x)); y .< median(y))
• iqr: x -> ((a, b) = quantile(vec(x), (1//4, 3//4)); a .<= x .<= b)

Each simulation study takes several hours to run. I'll update this issue with results as I complete the studies.