Improvements for Papers100m single gpu and single node multi gpu examples (Cugraph, GATConv, better default hyperparams, eval on all ranks)

[puririshi98]

Introduces improvements to NVIDIA GPU examples:

• Adds support for cuGraph data loading
• Adds GAT in addition to GCN
• Enables evaluation on all ranks
• Improves default hyperparams

~45% test acc w/ latest hyperparams on 8 gpus

latest cugraph single node multigpu works but drops degree 0 nodes in neighborloading causing lower test accuracy. @alexbarghi-nv will fix in 24.04 RAPIDS release. in general i think once cugraph 24.02 is out we can merge this PR and the test accuracy will automatically go up when using cugraph once users upgrade to 24.04 cugraph or newer

[puririshi98]

we get up to 5x speed ups with cugraph depending on the system used

[puririshi98]

in general GCN does better than GAT w/ about 45% test acc for GCN and slightly less for GAT. before the PR GCN was getting closer to 40% test acc

[puririshi98]

need @alexbarghi-nv to mirror my work on eval/testing on all ranks: https://github.com/pyg-team/pytorch_geometric/pull/8173/commits/c86f852644b3455bdaea6ee8cf8264eed6c5465a for cugraph then we will be ready to merge this

[puririshi98]

@akihironitta @rusty1s ready to merge

[akihironitta]

@puririshi98 I've just written the PR description, PTAL :)

[puririshi98]

needs a fix for cugraph multigpu from @alexbarghi-nv

[puririshi98]

okay @akihironitta @rusty1s, the current code in this PR only works with in 24.02 or newer cugraph so i will circle back for merging this once that is released publicly

[puririshi98]

@akihironitta @rusty1s anything needed to get this merged?

[puririshi98]

@rusty1s lmk if anything needed to merge

[rusty1s]

On it.