Add tests and documentation for "filter as a no-op" for each inference method

[SamDuffield]

It should be tested and documented for each method how to specify the parameters such that filter_prepare and filter_combine output an unmodified state.

This is useful because it simultaneously tests:

• The dynamics can be provided as the identity (useful if two observations are provided at the same model time but might be received at different real times in an online setting, or if we change to the convention of no intial observation).
• The likelihood/observation can be omitted and thus the filter step becomes a predict step (cuthbert doesn't distinguish between predict and update, only a unified filter step).

[SamDuffield]

Tests added in #186 but we should still document properly