Fix for divide by zero in psislw call from Pathfinder

[fonnesbeck]

This is intended to fix the divide by zero issue by wrapping logiw in xr.DataArray with dims="__sample__" before calling psislw. This is the documented-recommended input form per arviz's own psislw docstring.

Regression test added.

[jessegrabowski]

We need this to target the v0.11 branch on my fork of pymc-extras here . @ricardoV94 refused to help me figure out how to set up things like on pymc/pytensor to auto-target it, and I am playing helpless in order to complain for comedic effect.

This won't work on main because we're not cut over to arviz 1.0 yet, so CI will fail.

[aloctavodia]

az.psislw doesn't exist anymore on its own. Currently, that function should be used as an accessor (as already done in #668 ) or from the array_stats interface. I added a guard here arviz-devs/arviz-stats@bee3891. The test should pass then, as the _gpd_fit function should return np.nan, np.nan instead of getting into the divide by zero part.