opt for duplicate layer's implementations in GNN.jl and GNNLux.jl

Right now, Flux and Lux layers are just tiny wrappers are the forward pass implementation that lives in GNNlib.jl. While it is nice to share the code, I think that the price we pay in code readability is too high. 

Maybe we should just have duplicated implementations in GraphNeuralNetworks.jl and GNNLux.jl. 