[ENH] added a possibility for the network to append independent subnetworks into its structure, this functionality is then used in the ACSFnetwork to allow for dynamic construction of networks and possible generalizations of cluster interactions

[FIX] transfered the storage of activation values away from neuron object and put it into the neuralnetwork object

[FIX] Removed function load_from_file() and replaced with using a NeuralNetwork constructor instead.

[FIX] Removed explicit calls to init_from_file() and replaced with using a NeuralNetwork constructor instead.

[FIX] Randomization of parameters now seeded with milliseconds instead of seconds using std::chrono.

[NEW] Added method for evaluation the network on the whole DataSet, creating a new DataSet with predicted outputs.

[FIX] Rewritten serialization of ExprtkWrapper, so that Boost.serialization doesn't have to be friended anymore

[ENH] NeuralNetworkSum class now implemented using PIMPL to remove Boost::serialization header dependency induced by ExprtkWrapper.

[ENH] Non-parametric constructor in ExprtkWrapper made private, as its only purpose is serialization.

[ENH] fixed compilation, utilization of Salomon modules, inclusion of PETSc framework in the parallel version of levenberg-marquardt algorithm, minor fixes of the acsf example and run scripts