Automate choice of AD backend

[torfjelde]

We should have a method, say, choose_differentiation_backend(model), which simply benchmarks the model for different backends and returns the one that is the fastest (and correct).

[yebai]

I think this should be a functionality of DifferentiationInterface.jl, instead of Turing.jl

[penelopeysm]

With some of the refactoring I'm working on we could definitely put something like this in DynamicPPL 👀

[gdalle]

Check out the benchmarking utilities in DifferentiationInterfaceTest.jl ;)

https://juliadiff.org/DifferentiationInterface.jl/DifferentiationInterfaceTest/stable/tutorial/