WAIC/LOO only consider last chain in a Multitrace

[AustinRochford]

See this gist for a demonstration. Of course, the WAIC values across difference chains should not be too different, if the sampling has been effective. This may also affect other stat functions, I haven't checked yet.

@aseyboldt thoughts on what the correct behavior should be here?

[aseyboldt]

Hm. Intuitively I'd say we should probably merge the logp values from the traces, excluding tuning samles. It would be strange to use only one chain. That would though out a lot of data and the choice would be completely arbitrary.
So maybe check hasattr(trace, 'stat_names') and 'tune' in trace.stat_names or so and then filter out tuning samples.

[junpenglao]

Looks like a bug, we should just combine the traces when njobs>1.

[AustinRochford]

Hmm, I think the simplest solution here is, when iterating over a multi trace, to iterate over all chains, not just the last chain, as is currently done.

[junpenglao]

Is that how https://github.com/stan-dev/loo compute the loo/waic? Is it better to just combine all the traces together?