Make ess_rhat return NaN instead of erroring

[sethaxen]

Currently MCMCChains automatically displays our diagnostics for Chains. But for chains with less than 10 draws, each diagnostic that calls ess_rhat will error at the following line due to splitting:

MCMCDiagnosticTools.jl/src/ess_rhat.jl

Line 465 in 7d78e5d

throw(ArgumentError("number of draws after splitting must >4 but is $niter."))

This erroring is a nuisance to handle in either PosteriorStats or MCMCChains. I suggest we instead return NaNs in such cases and raise an informative warning about why ess_rhat failed. This would probably resolve TuringLang/MCMCChains.jl#447 and TuringLang/MCMCChains.jl#410

[devmotion]

We changed it to an error in #64 but I agree that probably NaN (+ @warn?) would be easier to handle for downstream packages.