Save data as attribute on fit objects

[WardBrian]

This was requested by @OriolAbril for better integration with Arviz. I think the main issue will be when we're creating a fit object from files, the data will no longer be around (unless we start saving it, which would fit with the ethos of #851 but could lead to unnecessary duplication of possibly-quite-large files)

[amas0]

To clarify, this is to save off whatever is passed to fit = model.<sampling_method>(data=data) as accessible with fit.data?

I don't think this would be too unreasonable? In the most common path of using sample with default settings, the output CSVs are probably much larger than the input data. With a minimum of 7 columns and default 1000 draws with 4 chains, then input data would need to be on the order of tens of thousands of numbers to approach the same output size. It's not nothing, of course, but I think it would rare for input data to be particularly large.

If we were to do something like this, I think the most sensible approach would be to save off the serialized input JSON (which we already construct in memory iirc) to the same output dir. It would be easy to make optional.

[WardBrian]

yep, exactly that. My understanding is that this would allow arviz to be more intelligent about placing items into the different groups of data/posterior/predictive when loading from cmdstanpy