Implement a GNN PPO for ray-rllib

[nhuet]

We follow the same guidelines as for the sb3 wrapper:

• GNN based on pytorch-geometric
• Feature extraction via GNN + reduction layer to a fixed number of
  feature
• Observation = Graph or dict whose values contains at least one Graph
• Action masks are taken into account if available
• User must use GraphPPO instead of PPO as algorithm: GraphPPO overrides
  PPO to change the way obs is converted to pytorch format

Worth noticing:

• We use the old api stack as the RLlib wrapper is currently using it
• For graph observations, the model is gnn extractor followed by a FullyConnectedNetwork
• For dict of graphs (and other) observations, the model is
  • preprocess obs by using gnn features extractor for graph components
  • apply to the prepreocessed obs a ComplexInputNetwork
• action masking is automatically activated according to domain class
  (not UnrestrictedActions) and algo class, as it was already coded in
  RayRLlib wrapper. The algo to be used is still GraphPPO as masking is
  managed by a custom model at RayRLlib wrapper level.