More flexible information for Runge-Kutta starter

[pnkraemer]

Currently, the Runge-Kutta starter uses jax.experimental.ode.odeint to generate data.
For stiff ODEs, this does not work.

Instead, I propose that we leave the data-generation to the user. This enables everyone to use jax.experimental, diffrax, scipy, or whatever to compute the first num_derivatives+1 observations. Maybe there are some "real" observations instead, and we may want to use those.

As a side result, this simplifies the code in runge_kutta_starter_fn massively.

It may be advisable to offer a helper function that preserves at least the current way.