Flatten/Destruct/Reconstruct functions in the package.

[ZeyuanLee]

In order to solve the system for nonlinear integrator, a destruct function and reconstruct function would be quite useful. The destruct should have two return values : the flattened parameters in the network, (NamedTuple{...} into a 1d vector), and a destruct function to refill the vector into network.

A generalized idea is to specify the degree of freedoms in a nonlinear expression (not just neural network), and flatten them into a vector, and meanwhile a reconstruct function.

Not quite sure whether these functions should be in SymbolicNeuralNetworks.jl or AbstractNeuralNetworks.jl

[michakraus]

This is a recurring issue. It should be implemented in AbstractNeuralNetworks.jl for NeuralNetworkParameters, possibly using ParameterHandling.jl.