More rigid radial cutoff parameters in EGNN.

[rousseab]

I fell in the trap of specifying radial_cutoff and edges inconsistently. Since the default value of edges is fully_connected, specifying only radial_cutoff does nothing and a fully connected calculation takes place.

To avoid giving the user the ability to shoot themselves in the foot (as I did), here I implement a consistency check on these two inputs.

I also corrected a "bug" in the test suite: the EGNN model should not be equivariant to the octahedral point group for arbitrary basis vectors; it should only be equivariant for a cubic unit cell. It is weird that the tests passed with a fully connected graph: I suppose that the lattice vectors don't matter in that case...