Implement summarystats for InferenceObjects types

[sethaxen]

This PR adds a StatsBase.summarystats method for Dataset and InferenceData inputs (EDIT: and also new utilities for nicely showing tables)

Internally, this calls several stats and diagnostics functions on the entire input Dataset and then concatenates them along a newly created :metric dimension. This new Dataset may be returned, but by default, the Dataset is turned into an iterator over marginals, which is interpreted as a row table and reformatted to be a column table. This column table is lightly wrapped in a SummaryStats object, which implements the Tables interface. The benefits of this object over returning a DataFrame are:

• we can in principle dispatch on this object in the future
• we can intelligently choose the display precision, e.g. as with ModelComparisonResult and AbstractELPDResult, we use the MCSE of an estimate, if available, to choose the number of significant digits to show of the estimate itself.
• we can avoid the heavy DataFrames.jl dependency

EDIT: Because this PR adds a number of utilities for nicely showing tables, these are now used in the show methods of ModelComparisonResult and AbstractELPDResult. For the latter, I also reformatted the table to be more similar to that of ModelComparisonResult. Since changes to show methods should never be considered breaking changes, this applies here as well.

Future improvements

We can further support users providing their own stats funs to include/replace the ones used, but doing this well would involve adding a utility function for users to wrap a function that reduces to a scalar or a Tuple or NamedTuple of scalars so that it maps across all non-sample indices for all variables, producing a new Dataset, optionally with a new dimension. This would actually reduce a lot of code redundancy, so this utility should appear in InferenceObjects first.

Similarly, the flattened iterator will be useful for plotting and may need to be refactored to become an API utility as well. This could certainly be done in a more type-stable way.

Without any additional work, we could use Crayons to highlight bad ESS values and R-hat values. This would just require storing information about number of chains/draws in the SummaryStats object.

[sethaxen]

Without any additional work, we could use Crayons to highlight bad ESS values and R-hat values. This would just require storing information about number of chains/draws in the SummaryStats object.

I'm locally testing this, and I think it looks pretty good. It keeps with our trend in e.g. psis of using colors in the REPL to draw attention to diagnostics that look bad.

We could additionally add informative warnings, something like 10 params have ESS values lower than 100*nchains. and 5 params have R-hat values higher than 1.01.

[sethaxen]

I've opted for now to not do highlighting or warnings. This can be addressed in a future PR.

[sethaxen]

I've checked that the tables look good in the REPL, Documenter, jupyter, and Pluto.

using ArviZ, ArviZExampleData
idata = load_example_data("radon");
summarystats(idata)

The REPL and doctests use the plain-text show method. Here's the REPL version:

jupyter and Documenter use the HTML show method. Here's the jupyter version:

For better or for worse, when Pluto sees that an object implements the row table Tables interface, it ignores all show methods and instead uses its own semi-interactive table viewer: