Currently, sample_conditional_posterior and extract_components_from_idata return square brackets notations for the state coord, making it hard to select components when working with multivariate series, as showed in #544.

As discussed with @jessegrabowski and @OriolAbril, we should return an xarray.Dataset with one entry per observed state (e.g ['chirac2', 'sarkozy', 'hollande', 'macron', 'macron2']), thus enabling users to select the corresponding structural components with .sel. This also allows each observed state to have arbitrary model structure inside, which the current multivariate setup allows.