Replace StateSpaceVar.condition_on_qoi_observation

[pnkraemer]

Replace condition_on_qoi_observation(u, std) with e.g. observe_qoi which computes only the marginals and the backwards transitions. No data.

If desired, the correction can be computed with backward marginalisation.

The advantages of this split are:

• No argument duplication in log-likelihood computation: currently, the input (u) enters both the condition_on_* and the logpdf function, but only the latter is necessary for the log-likelihood computation
• (Related to the above) We an cache the output, which implies more efficiency
• Simplification of implementing linear operator observations (link: Linear operator observations in negative-marginal-likelihood #265). Implementing such observations becomes a one-liner in dense_output.py instead of a 10-liner in each of the state-space model implementations.