ADgradient should always cache/prepare #51
Description
Given that the API provides a vector length, "preparation" (allocation of buffers, caches, precomputations, everything that DifferentiationInterface.prepare_... methods do) just needs a type.
I am proposing that ADgradient always performs this preparation step, with the default type T = Float64. The user may call the function with other types, in which case the layer should just silently convert as needed, without warning.
Cf JuliaDiff/DifferentiationInterface.jl#859, but we can do this in this package if that is not supported.
The x = ... argument should be retired.