Add k_endog argument to structrual Components to enable multivariate structural models

[jessegrabowski]

Currently, we can only do univariate timeseries analysis with the structural framework. There's no reason for this -- the same framework can be extended to multiple time series. We would just need some extra API to support it. Basically:

• All components would need a k_endog and endog_names arguments, which would determine how many variables we're modeling
• When we call build, we need to gather up the set union of all endog_names. They don't have to all match, because we could have some components in one but not another
• We need some new logic to build the Z matrix to map the right hidden states to the right observations.

The biggest challenge would be if we want to allow latent state sharing. For example, I could imagine a model where we want several time series to share the same LevelTrend component. In this case, we might need an additional argument, like shared_endog_names or something? Then instead of copying + block-concatenating the relevant matrices, we map all the different observed timeseries to the same hidden states via the Z matrix.

Obviously this is all just off the top of my head, suggestions/criticisms highly welcome.