Automatically create graph models

[klamike]

Graph-based proxies are all the rage recently, but they can be a bit cumbersome to implement. We can alleviate that in L2O by extracting a graph representation automatically so users can easily build a GNN model. I imagine an API like:

import LearningToOptimize as L2O
import GraphNeuralNetworks as GNN

model::JuMP.Model = ...

encoder = L2O.graph_encoder(L2O.Bipartite, GNN.GATConv, model; out=128)

gnn = Flux.Chain(
  encoder,
  Dense(128,1)
)

https://github.com/lanl-ansi/MathProgIncidence.jl should be very helpful for this.