Add partial and complete missing emission support in LGSSM filtering

[hylkedonker]

This pull requests adds support for the inclusion of missing measurements/emissions/observations $y_t$ in the Kalman filter dynamax.linear_gaussian_ssm.lgssm_filter. This PR covers:

• Fully missing measurements. For example, $y_t = [y_{t1}, y_{t2}]^T$ where both elements are missing.
• Partial missing observations, where for example $y_{t1}$ is missing, but $y_{t2}$ is observed.
• Includes comprehensive unit tests.

This PR partially addresses issues #301, #330, #399

Implementation details
Let $P_{t+}$ by the operator that projects onto the observed values of $y_t$. Then the update step of the Kalman filter is modified by replacing:

• $H_t \rightarrow P_{t+} H_t$,
• $R_t \rightarrow P_{t+} R_t P_{t+}$.

Mathematical details are provided in the comment of issue #399.

In addition, the log-likelihood term is adjusted accordingly, e.g., $ \int \mathrm{d} y_{t1} p(y_{t1}, y_{t2}|y_{1:t-1})= p(y_{t2}|y_{1:t-1})$ while keeping the shape of the covariance matrices static (to prevent recompilation).

Scope
Not yet included are the corresponding implementations in:

• parallel_lgssm_filter
• lgssm_info_filter

Which I consider out of scope of this PR.