Closes #545

Currently extract_components_from_idata returns square brackets notations for the state coord, making it hard to select components when working with multivariate series. Things like:

state = [
            "trend[level[gdp]]",
            "trend[trend[gdp]]",
            "trend[level[unemployment]]",
            "trend[trend[unemployment]]",
            "ar[gdp]",
            "ar[unemployment]",
]

This PR adds a restructure argument to extract_components_from_idata (default False for backwards compatibility). When True, it will restructure the state coordinates as a multi-index for easier component selection, thus enabling selections like idata.sel(component='level'), idata.sel(observed='gdp'), or even idata.sel(component='level', observed='gdp').
Again, this is especially useful for multivariate models with multiple observed states.

More precisely, the state dimension is broken down into two new ones, component and observed, whose coordinates will be [('level', 'gdp'), ('trend', 'gdp'), ('ar', 'gdp')], [('level', 'unemployment'), ('trend', 'unemployment'), ('ar', 'unemployment')] .

This also allows each observed state to have arbitrary model structure inside, which the current multivariate setup allows.

NB: This PR is a first pass, and in no way exhaustive -- we probably need to expand to more complex cases, that users will surface up. But at least it gets the ball rolling and should be self-sufficient to already merge.