[GraphBolt][CUDA] Showcase GPUCachedFeature usage in an example.

[mfbalin]

No description provided.

[mfbalin]

@frozenbugs Do you think anybody would like to work on adding such an example? We could showcase it after a larger dataset is added to GraphBolt in as part of #6909. We can then showcase the speedups hopefully, however, to show speedups, the feature copy needs to dominate the overall training runtime. But with #6954, the feature copy can be fully overlapped so it is going to be a bit harder to show speedups after it is merged.

When we have a large dataset such as ogbn-papers100M or maybe the mag240M hetero-rgcn.py example, we need to wrap its main TorchBasedFeature with a GPUCachedFeature in #6939.