simplify mcse_elpd using log-normal approximation

[avehtari]

Years ago I made mcse_elpd using a simple quadrature, but now switched to use a more efficient log-normal approximation. Given the mean and variance in linear scale, and knowing that the logarithm is closer to normal, we approximate the uncertainty in linear scale with log-normal and compute the corresponding normal distribution parameters in log scale using the equations shown, e.g., at https://en.wikipedia.org/wiki/Log-normal_distribution#Arithmetic_moments

The change makes relative change of 0.01%-0.1% compared to stored reference values and thus some reference rds files were updated.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (ce0f540) 92.45% compared to head (6d348a1) 92.44%.

❗ Current head 6d348a1 differs from pull request most recent head 473e107. Consider uploading reports for the commit 473e107 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #246      +/-   ##
==========================================
- Coverage   92.45%   92.44%   -0.02%     
==========================================
  Files          31       31              
  Lines        2835     2831       -4     
==========================================
- Hits         2621     2617       -4     
  Misses        214      214              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jgabry]

Looks good. I'll include this in the release.