Return posterior predictive samples from all chains in ModelBuilder

[mbjoseph]

This fixes a bug where only values from one chain were returned. It also refactors the prediction logic to reduce duplication. I've used arviz.extract() to get the posterior predictive samples as hinted by @twiecki in #139, removing the _extract_samples() method from the model builder class.

Behavior change proposal

This PR also includes a proposed behavior change that makes the output of predict_posterior() consistent in type and shape with the output of pymc.sample_posterior_predictive().

Namely, predict_posterior() would here return an xarray Dataset from the posterior predictive distribution, rather than a dictionary of numpy arrays. This is an admittedly opinionated change in behavior, though it does also streamline the code. From my perspective, retaining the metadata related to chains, draws, and other attributes is useful for three key reasons:

1. downstream processing for posterior predictive checks, enabling easier alignment of chains and draws with other pushforward quantities, and
2. tracking the provenance of predictions (e.g., creation time and package versions attributes) in a production environment, and
3. consistency with the output from pymc.sample_posterior_predictive(), which could make behavior easier to predict for users, and allow greater interoperability with other functions that process and plot posterior predictive samples

That said, I'm very open to other perspectives on how to bundle up posterior predictive samples.

[michaelraczycki]

Looks good, for the future reference please try not to just switch places of functions if it's not needed, as it makes the code changes less readable for PR

[mbjoseph]

Thanks for the review @michaelraczycki. I've added combined as a named argument, tested behavior of combined=True and combined=False, and restored the original method order in 02faac7.

[michaelraczycki]

Looks good to me

[twiecki]

Thanks @mbjoseph and for the review @michaelraczycki!