[integration/peerlab] Bugfix/torch array overlow

[Tobiaspk]

Closes #44 . Subsetting a graph with over 2.1B edges is not supported in torch 2.5.1 (which is often the version it's resolved to). Instead, use indices instead of boolean arrays to subset the graph. These can be obtained using chunks.

Also improve logging logic.

[Tobiaspk]

Attached is the script used to test if the new chunking logic yields identical results to the previous implementatino.

test_new_subsetting_logic.py

Outputs:

Loading HeteroData and Tiles...
#tiles: 13
#nodes: 0.3M
#edges: 1.5M
Splitting mask of size 280378 into 1402 chunks...
Checking if new and previous subsets are equal...
Node type tx matches.
Node type bd matches.

@andrewmoorman FYI, will merge tomorrow if you see no issues. :)

[Tobiaspk]

Prediction graph now works. While performance for predicting is very slow now, this fix solved the issue.