Add cuML KNN for graph construction, better handling of non-PBC systems

[vsimkus]

• cuML for KNN graph construction
• Determining if a system is periodic using ase.Atoms.pbc and ase.Atoms.cell rather than just ase.Atoms.cell.

I've added a note to the README regarding optionally installing cuML.

I've tried adding cuml as an optional dependency in poetry such that users can pip install orb-models[cuml-cu11] or pip install orb-materials[cuml-cu12] depending on their system. But poetry tries to resolve dependencies between the two optional packages cuml-cu11 (for cuda 11) and cuml-cu12 (for cuda 12) and fails since they require different sub-dependencies. I don't think there is currently a way to specify mutually-exclusive optional dependencies in pip/poetry right now. So it's probably best to just have it in the README.

[vsimkus]

On that note, should we perhaps set the default edge_method in Calculator to None:

orb-models/orb_models/forcefield/calculator.py

Line 21 in 385e091

edge_method: Optional[EdgeCreationMethod] = "knn_scipy",

such that the default method would be chosen dynamically:

orb-models/orb_models/forcefield/featurization_utilities.py

Lines 38 to 54 in 385e091

	def get_default_edge_method(
	device: torch.device, num_atoms: int, is_periodic: bool
	) -> EdgeCreationMethod:
	"""Get the default edge method for a given device and number of atoms."""
	if device.type != "cpu":
	if (
	cuml is None
	or (is_periodic and num_atoms < 5_000)
	or (not is_periodic and num_atoms < 30_000)
	):
	edge_method = "knn_brute_force"
	else:
	edge_method = "knn_cuml_rbc"
	else:
	edge_method = "knn_scipy"
	assert edge_method in typing.get_args(EdgeCreationMethod)
	return edge_method # type: ignore

[DeNeutoy]

Super, LGTM!