Add n_eff and r_eff to kfold diagnostics

[florence-bockting]

Description

This PR adds n_eff and r_eff to the kfold() diagnostics slot (which currently only incorporates pareto_k). The computation logic is as follows:

1. unnormalized log weights: $\log(w_{si}) = - \log[p(y_i \mid \theta_s)]$
2. normalized weights: $\log(\tilde w_{si}) = \log(w_{si}) - \log[\sum^S_{s'=1} w_{s'i}] \rightarrow \tilde w_{si} = \exp[\log(\tilde w_{si})]$
3. effective sample size: $n_{eff,i}= 1/\sum_{s=1}^{S} \tilde w_{si}^2$

whereby $r_{eff,i} = 1$.

TODO

• add computation for n_eff and r_eff
• add n_eff and r_eff to kfold diagnostics output
• add unittests for updated output structure

[florence-bockting]

@avehtari Can you check whether the computation is correct? Additionally, I am uncertain about r_eff. It is currently hard coded at 1, not sure whether this is appropriate.

[avehtari]

If the posterior inference has passed convergence checks, r_eff tends to not have much influence to MCSE, and for simplicity we decided to switch using default 1. loo() output includes now line telling r_eff range, so that it is possible to see whether 1 or comouted values are used. I'll check the rest later

[florence-bockting]

@avehtari Can you please check the final computation of r_eff, n_eff and k_threshold. Thank you.