Split discrete and continuous Kalman filter implementations

[pnkraemer]

In a Nutshell

Splits the current Kalman implementation into a discrete and a continuous part. The DiscreteKalman does Gaussian filtering and smoothing for discrete-discrete state-space models. The ContinuousKalman does the same for continuous-discrete state-space models. The goal of this PR is to make another step towards easier-to-maintain filtering and smoothing code.
(This PR also does some first mini-steps in simplifying the filter() implementation in the discrete object.)

Detailed Description

The filter() and smooth() functions in the Kalman filter currently carry around a lot of dead weights for discrete models. Most notably, a dt variable, which does not make sense. (Also, the t variable is pointless, but this can only be refactored as a part of a bigger effort; detailed in #627.)
Soon, when there are suitable priors/problem classes for discrete vs continuous state-space models (see also #646), and when approximate Kalman filters are refactored, the gains obtained through the separation introduced here will be even more evident.

Some further changes:

• Iterated filtsmooth in Kalman was not used anymore, so I deleted it. (It is superseded by filtsmooth.gaussian.optim)
• discrete-discrete iterated Kalman smoothing did not have a test case, which I added.

[codecov]

Codecov Report

Merging #645 (2408077) into main (3e2e59b) will increase coverage by 0.13%.
The diff coverage is 100.00%.

❗ Current head 2408077 differs from pull request most recent head c64d6a6. Consider uploading reports for the commit c64d6a6 to get more accurate results

@@            Coverage Diff             @@
##             main     #645      +/-   ##
==========================================
+ Coverage   90.12%   90.26%   +0.13%     
==========================================
  Files         194      194              
  Lines        7353     7343      -10     
  Branches     1163     1160       -3     
==========================================
+ Hits         6627     6628       +1     
+ Misses        486      470      -16     
- Partials      240      245       +5     

Impacted Files	Coverage Δ
.../odefilter/init_routines/_non_probabilistic_fit.py	100.00% <100.00%> (ø)
src/probnum/filtsmooth/_kalman_filter_smoother.py	100.00% <100.00%> (ø)
src/probnum/filtsmooth/gaussian/__init__.py	100.00% <100.00%> (ø)
src/probnum/filtsmooth/gaussian/_kalman.py	96.72% <100.00%> (+28.46%)	⬆️
src/probnum/quad/solvers/bayesian_quadrature.py	86.04% <0.00%> (-7.29%)	⬇️
src/probnum/quad/_bayesquad.py	94.44% <0.00%> (-5.56%)	⬇️
...bnum/quad/solvers/belief_updates/_belief_update.py	96.96% <0.00%> (-0.09%)	⬇️
src/probnum/typing.py	100.00% <0.00%> (ø)
src/probnum/quad/__init__.py	100.00% <0.00%> (ø)
... and 17 more

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[pnkraemer]

This wrapper around the state generator works with args and kwargs for the following reason:

Soon, the regression problems might change types (discrete vs continuous), and the filter() and iterated_filtsmooth_generator() signature adapts accordingly. But this function here (and the filtsmooth() style methods below) remains the same. args and kwargs lead to a smaller, more meaningful diff in the future.