[Future] Use DifferentiationInterface as primary interface to AD-backends

[oschulz]

For the (probably) next breaking release of AutoDiffOperators

• Fully depend on DifferentiationInterface and drop (almost) all AutoDiffOperators-native backend-related code.

• Remove all support for "structured" arguments and focus AutoDiffOperators on vector arguments and AD-based multiplicative operators that work on them.

Replaces #17.

[oschulz]

With #31, this is finally completed. AutoDiffOperators uses DifferentiationInterface throughout now. DI-prep used used only for Jabian operators, so far, and uses a thread-safe caching for DI prep objects. ToDo: Also use prep for gradients.