single node multiGPU node prediction

[puririshi98]

closing old PR to fix merge conflict: https://github.com/snap-stanford/relbench/pull/125

[puririshi98]

@kexinhuang12345 @weihua916 @joshr17 reviews welcome. please lmk if anything else needed to merge

[rishabh-ranjan]

Hi @puririshi98, thanks for this script. However, I am not sure this script belongs in the RelBench repo.

1. The multi-gpu part has nothing to do with RelBench functionalities and everything to do with GNN training (please correct me if I am wrong). As such, a multi-gpu example in pytorch-geometric seems more appropriate.
2. We currently don't have datasets/tasks with scale that requires this. But there is indeed interest in working on a large-scale version of the benchmark, which we can chat about separately.