Consider Supporting rstanarm's Variational Inference algorithms

[andytimm]

The idea here would be that rstanarm's variational inference (algorithm = "meanfield" and algorithm = "fullrank") implementations would be neat to support, and aren't bad at all to integrate.

So I have a portion of this working locally already, and am happy to test and put in a PR, but since my solution for implementing this feels moderately hacky, opening an issue to see if this is acceptable, or if you have a better suggestion.

Briefly, the idea is that chains don't exist for VI. The simple solution I've implemented here essentially just gets tidy_draws to treat this type of stanreg object like a more typical mcmc stanreg object with 1 chain. It's a little inaccurate to label this way, but sidesteps what I imagine would be a bunch of issues with not having a .chain column.

If that sounds worth bringing into the package, or you have a more elegant solution, let me know. Thanks!