Better performance guidelines and support for user

[torfjelde]

We currently have some tips for how to improve performance of a Turing.jl model.
However, these are more "tips & tricks" rather than a good introduction on how to a) benchmark and profile a model, b) which AD backend to choose, and c) how to write models in a way to make them more efficient.

So, IMO, there are a few things we should do:

• Write a proper tutorial one how to speed up a model. This includes the three listed properties above. Ref: Tutorial on Turing.jl and performance Turing.jl#2415.
• [ ] Offer functionality to automatically make some choices for the user. Ref: Automate choice of AD backend Turing.jl#2417
• Provide suggestions programmatically either "always" or behind an explicit function, e.g. performance_hints(model). Ref: Performance hints Turing.jl#2416