Migrate to torch-cluster

[JPXKQX]

This PR improves the performance of creating a graph:

• Update the KNNEdges & CutOffEdges to rely on torch-cluster. This method also supports moving the tensors to the GPU.
• EdgeAttributeBuilder's are now implemented as message passing graphs. This method supports moving the tensors to the GPU.

The performance improvement for generating a graph n320 (~500k nodes) -> o96 (~40k nodes) -> n320 is:

version	Device	CutOffEdges(cutoff_factor=0.6)	CutOffEdges attribute	KNNEdges(num_nearest_neighbors=3)	KNNEdges attributes
v3.0.0	CPU	2.25 s	0.55 s	11.83 s	1.21 s
torch-support	CPU	0.45 s	0.31 s	0.30 s	0.54 s
torch-support.	1 GPU	0.19 s	0.14 s	0.15 s	0.00 s

The main difference is that the Haversine distance is not supported in the torch-cluster. To solve this problem, we transform the 2D coordinates into 3D coordinates (sphere) before calculating the Euclidean distance. The edge direction calculation has also been refactored to use torch.tensor's instead of np.array's & scipy (using Rodrigues' rotation formula).