Add change of default AD backend

[ocots]

#25

[gdalle]

My apologies but that's extremely ugly 🤣 why don't you just create a backend kwarg that gets passed all the way down from the user-facing functions / structs to the functions actually doing the differentiation?

[gdalle]

Not to mention that you might need a combination of backends: you don't necessarily want to use the same backend for gradient and jacobian for example