Improve performance of bayesian model estimation

[p-gw]

Before the first release we should try to make models as performant as possible.

[t-alfers]

Turing.setrdcache(true) and Turing.setadbackend(:reversediff) is not provided anymore. Any reasons? Without these calls the model estimation is extremely slow. At least we should provide information for the user to load Turing and call these two functions after loading RaschModels.jl.

[t-alfers]

At least for the RaschModel with ReverseDiff and caching enabled, it looks quite nice, actually.

[p-gw]

Turing.setrdcache(true) and Turing.setadbackend(:reversediff) is not provided anymore. Any reasons?

No, probably just slipped through when I refactored fit... Of course this needs to be put back in.

At least for the RaschModel with ReverseDiff and caching enabled, it looks quite nice, actually.

Yes, the rasch model I think is pretty good already, but I have the feeling that the polytomous models can be improved