Asynchronous methods on feature/graph stores

[aaronzo]

The addition of feature/graph stores offloads sampling and getting from the machine running PyG to a remote store, e.g. a graph database or other. The async framework in python is suited for IO operations over network requests - should the remote backends support this? This might also need a DataLoader which supports asynchronous partition pre-fetching.

Currently, setting the num_workers > 0 on a torch.utils.DataLoader creates a whole new python process for doing work, and can have memory issues as documented here: pytorch/pytorch#13246 (comment). In the graph/feature store remote case this is overkill; the worker processes are just going to be blocked waiting for synchronous requests to be responded to. Instead, asynchronous requests from the main process could do this job, with little to no performance impact.