Support missing (`nan`) emissions in `LinearGaussianConjugateSSM.fit_blocked_gibbs`

Hi,

First off, thanks for building this awesome package. Well done!
I work in the hospital and I am interested in fitting a Kalman filter on sparsely measured blood levels where most observations are missing. Am I correct in assuming that missing values are currently not supported by LinearGaussianConjugateSSM?
If so, in theory this is not so difficult to implement in the fit_blocked_gibbs if an entire emission is missing, I think.

1. Update forward filtering backwards sampling algorithm with the missing observations integrated out.
2. Update the parameter sampling. Integrating out the missing observations would amount to removing the control-emission pairs from the emission summary statistics where the emission value is missing. More complex would be when part of the emission is missing (giving marginal emissions). That probably requires a generalization of the normal-inverse Wishart distribution. See also this paper: Missing observation analysis for matrix-variate time series data.

If you're interested, I can give the complete missing observation a try. Would you accept such a pull request?

Kind regards,

Hylke

Start at LinearGaussianConjugateSSM.fit_blocked_gibbs and trace the forward filtering backward sampling and parameter-sampling paths. Done would mean supporting entirely missing emissions by integrating them out and excluding their control-emission pairs from emission summary statistics; partially missing emissions are described as a more complex follow-up.