Autodiff gradients for parameters error

[lindnemi]

https://github.com/PIK-ICoN/NetworkDynamics.jl/blob/master/test/autodiff_test.jl

I don't know why ReverseDiff started throwing an error as well, but it does now. The problem seems to be for ForwardDiff, as well as for ReverseDiff, that the Tracked resp. Dual numbers can't be stored in our internal caches because of type mismatch.

An idea would be to dispatch on parameter type in prep_gd, and change the types of the buffers. However, the tuple syntax means that we would have to dispatch on many different possibilites. Hence, fixing this should be combined with #34.

[lindnemi]

Forget what I said. The problem is solved by initializing the output array with the correct type (matching the type of the parameters). All autodiff tests pass now.