OGBMAG240m example

[puririshi98]

Epoch: 20, Step: 1000, Loss: 0.8472, Train Acc: 72.31%, Most Recent Step Time: 0.1662s
Epoch: 20, Loss: 0.8822, Train Acc:72.31%, Average Step Time: 0.1624s
Validation Accuracy: 68.7001%
Test Accuracy: 68.6972%

PR finally ready, test accuracy is on par with claimed accuracies in OGB repo

[puririshi98]

@akihironitta lmk if theres anything else needed to merge this, we have been using this internally and in our containers for some time now, it would be great for all PyG users to have access.
snap-stanford/ogb#449
this also needs merging

[akihironitta]

@puririshi98 It would also be nice to add a quick docstring of the model args.

Also, would it make sense to have this script in examples/multi_gpu/ and add examples/multi_gpu/README.md briefly describing what each of the multi-gpu scripts (that you recently added) does and/or how each example differs from each other?

[puririshi98]

@akihironitta addressd, ur comments, lmk if you have any further review requests

[puririshi98]

@rusty1s @akihironitta lmk if there is anything else i would need to get this merge ready

[akihironitta]

Looks great! I'll just keep my review a "comment" as I'd like to run this in my env, too :) Feel free to ask Matthias to merge if I'm too slow ;)

[puririshi98]

@akihironitta thanks for the review, no rush, please merge once your ready.

[rusty1s]

Nice. Note that you also need to call acc.reset() in order to avoid sharing metrics between training/validation.