make E_loo Pareto-k diagnostic more robust

[avehtari]

Fixes #250 (although the reason was not posterior version)

E_loo was giving warnings and errors in certain cases with binary x. If h(theta) is a step function and h is 0/1 binary, then when computing Pareto-k the one of the tails is likely to be just 0's. In such cases, using the same argument as for type="quantile", we can return the Pareto-k just for r. To make the function to complain less, the same approach is now used for constant, NA, NaN, and infinite x. Added also tests that these don't cause warnings.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 85.71429% with 1 lines in your changes are missing coverage. Please review.

Project coverage is 92.76%. Comparing base (45c261e) to head (42f24bd).

❗ Current head 42f24bd differs from pull request most recent head be0266c. Consider uploading reports for the commit be0266c to get more accurate results

Files	Patch %	Lines
R/E_loo.R	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #251      +/-   ##
==========================================
- Coverage   92.78%   92.76%   -0.03%     
==========================================
  Files          31       31              
  Lines        2829     2834       +5     
==========================================
+ Hits         2625     2629       +4     
- Misses        204      205       +1     

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

[jgabry]

Looks good. I fixed a typo in a comment so the unit tests are running again but this is ready to merge when they finish running.