Simplify init_prepare

[Sahel13]

As discussed in #266

@SamDuffield Thoughts on the changes to init_prepare for Kalman?

[SamDuffield]

My initial thought is I'm sure I like this as it means the user would have to rebuild the filter if m0 and chol_P0 changed, but I guess that's ok (it's what we do anyway in the EM example)

I like that we remove model_inputs from init_prepare as it makes it clear the user should never expect a non-dummy model_inputs at the initial state.

So overall happy to go with this, although maybe good to confirm with @AdrienCorenflos too as it's an API breaking change