Allowing alternative InferenceData/Dataset storage

Currently all `Dataset`s wrap a `Dimstack`, which in turn wraps a set of `NamedTuple`s. Likewise, `InferenceData` wraps a `NamedTuple` of `Dataset`s. As noted in #15, this makes everything but the array values themselves immutable.

An alternative is to decouple the interface for `Dataset` and `InferenceData` from the storage. e.g. we could define `NamedTupleStore`, `DictStore`, `NCDatasetStore`, etc and implement the `InferenceData`/`Dataset` interfaces for each of these. This would allow users to strongly type everything with `NamedTupleStore` if they want to, or they could use a `DictStore` to have much more dynamic access. Having a `NCDatasetStore` would allow users to open a NetCDF file as an `InferenceData` and even incrementally write to such an `InferenceData`, having the NetCDF file automatically updated.

With a well-designed API for stores, this would actually look a lot like the InferenceData API proposed in https://github.com/arviz-devs/ArviZ.jl/issues/154. Namely, the store API could be implemented for a type like `MCMCChains.Chains`, and then one could call `InferenceData(chains)` to (inefficiently) view it as an `InferenceData`. When efficiency is needed, one just converts to one of the native stores.

I started work on a small prototype of this. It's a fair amount of surgery to the existing code and increases the code complexity. Though it will be more work to do it then than now, I think it's better to hold off on this until later.