Add warning for log scale default in compare/loo/waic functions

[AlexAndorra]

Description

This PR adds a reminder about log scale default and how it should be interpreted (higher is better) for az.waic and az.loo. This merely mirrors what we already say in the doc.

For az.compare I added a UserWarning following the change to log scale in 0.7.0. This adresses this issue, and the warning will be temporary, as discussed in the issue.

In passing, I trailed all the white spaces before colons in the docstrings of stats.py. This messes up the formatting on the website.

I'm here for any change, and thanks in advance for the review 🖖

[OriolAbril]

I personally like the space before the colon, and it looks like it is common formatting (see pandas.read_csv), but I would not oppose to changing the format

[AlexAndorra]

Yeah, but weirdly it also take the argument's type when double-clicking it. When you remove the white space, it just selects the parameter, as it should -- don't know why 🤷

Thanks for the review Oriol, I updated in light of your comments.

[OriolAbril]

Yeah, but weirdly it also take the argument's type when double-clicking it. When you remove the white space, it just selects the parameter, as it should -- don't know why shrug

Ohh, true, I hadn't noticed, and it looks like it happens with any documentation with space before the colon 🤔

[AlexAndorra]

Yeah that's weird 😅

Do we still need assert loo_data == loo_pointwise[: len(loo_data)] in test_loo_print? It's this line that messes up the tests, because it's no longer the case.

[ahartikainen]

I think we don't need it. What is it now?

[AlexAndorra]

Now it includes the warning I introduced in this PR:

Computed from 2000 by 8 log-likelihood matrix
         Estimate       SE
elpd_loo   -30.81     1.43
p_loo        0.95        -
The scale is now log by default. Use 'scale' argument if you rely on a specific value. 
A higher log-score (or a lower deviance) indicates a model with better predictive 
accuracy.

when default,

Computed from 2000 by 8 log-likelihood matrix
         Estimate       SE
elpd_loo   -30.81     1.43
p_loo        0.95        -
------
Pareto k diagnostic values:
                         Count   Pct.
(-Inf, 0.5]   (good)        6   75.0%
 (0.5, 0.7]   (ok)          2   25.0%
   (0.7, 1]   (bad)         0    0.0%
   (1, Inf)   (very bad)    0    0.0%
The scale is now log by default. Use 'scale' argument if you rely on a specific value. 
A higher log-score (or a lower deviance) indicates a model with better predictive 
accuracy.

when pointwise.
Which explains why the test is failing.

[OriolAbril]

It can be removed, black is also complaining, make sure to run black right before commiting

[AlexAndorra]

Done! Thanks guys 👌

[codecov]

Codecov Report

Merging #1150 into master will increase coverage by 0.13%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1150      +/-   ##
==========================================
+ Coverage   92.99%   93.12%   +0.13%     
==========================================
  Files          94       94              
  Lines        9242     9286      +44     
==========================================
+ Hits         8595     8648      +53     
+ Misses        647      638       -9     

Impacted Files	Coverage Δ
arviz/stats/stats.py	96.33% <100.00%> (+<0.01%)	⬆️
arviz/stats/stats_utils.py	91.72% <100.00%> (+0.05%)	⬆️
arviz/data/io_numpyro.py	95.23% <0.00%> (+3.64%)	⬆️
arviz/data/io_dict.py	89.89% <0.00%> (+7.24%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d1e1a2a...1272225. Read the comment docs.

[OriolAbril]

very minor comments. it should also be added to the changelog and will be ready to merge

[AlexAndorra]

Thanks Oriol! It's now updated -- and I had already added it to the changelog 👌