Be able to define define the node features at the level of the network

Context: I've found to get sensibly better result using x, y, z for the GNN.

If we want to use r, \phi, z for the embedding and x, y, z for the GNN, this MR makes it possible.

Basically, instead of using

gnn:
  feature_indices: 3

You can use

gnn:
  feature_names: ["un_x", "un_y", "un_z"]
  feature_means: [0.0, 0.0, 281.0]
  feature_scales: [14.6, 14.6, 287.0]

(un_x, un_y, and un_z are defined at the processing stage).

You can still use the old version.

This is

• much clearer
• much more flexible
• needed

This MR also fixes bugs here and there, and edit the configuration files to remove the old use_xyz variable.