Add support for inhomogeneous parameters in LinearGaussianConjugateSSM.fit_blocked_gibbs

[hylkedonker]

This PR aims to address #402. In brief, add support for inhomogeneous (i.e., time-varying) parameter Gibbs sampling in the linear Gaussian conjugate state space model LinearGaussianConjugateSSM.
The primary code change in LinearGaussianConjugateSSM.fit_blocked_gibbs is this:

• Compute the summary statistics per time point $t$.
• Sample the the parameters
  • Inhomogeneous model: Sample a separate $[F_t, B_t, b_t]^T$ and $[H_t, D_t, d_t]^T$ from a multivariate normal inverse Wishart (MNIW) posterior for each time point $t$.
  • Homogeneous (steady-state) model: Aggretate the summary statistics from all time points, then sample $[F, B, b]^T$ and $[H, D, d]^T$ from the MNIW.
• Includes a (fairly trivial) unit test that checks the sampled shapes. (A more comprehensive Markov chain Monte Carlo test would involve simulation based calibration, but these take too long to be included in a test suite.)

Scope:

• Currently, uses the same multivariate normal inverse Wishart prior for all time points.
• Either considers all dynamics and emissions weights and biases as time-dependent, or all time-independent. For instance, this PR doesn't yet provide support for time-independent emissions with time-dependent dynamics (or vice versa).
• Algorithms other than fit_blocked_gibbs -- e.g., EM -- are out of scope of this PR.

Let me know if this looks good, or if you require any modifications.