[integration/peerlab] Optimisation/tiling

[Tobiaspk]

Closes #49

This implements a CSR-like subsetting approach as highlighted in #49. It consists of 2 majors updates:

1. Sort all edges during graph construction by src.
2. Replace data.subgraph with a custom subgraphing implementation that is very similar to the torch_geometric source code, but uses the proposed subsetting logic.

Tested this by comparing the subsets of the old and new version on 100 randomly selected prediction tiles of the public Atera Breast Cancer dataset. Results

• Number of nodes and edges are always identical
• The prediction mask sometimes changes in up to 0.3%. This could be caused by replacing the buffer operation with a simple margin expansion of the tile.

Performance improvement in this dataset is 8.5x.