Allow S3 file source for distributed training and partitioning

## 🚀 Feature
Currently, the distributed partitioning tools assume the existence of a shared filesystem in which all necessary files are available for each node in the clusters.

It would be easier for the user side if they were able to provide a source location on S3 and then individual workers downloaded the files they needed, instead of assuming the presence of a shared FS.

## Motivation

Setting up and maintaining a shared FS is an additional friction point for someone looking to try out distributed training in an enterprise setting.

## Alternatives

Our (GraphStorm) current solution is to download these files in our wrapper scripts, but this creates entanglement with DGL.

## Pitch

Add the ability for users to provide an S3 prefix under which data are arranged in a manner described in https://docs.dgl.ai/guide/distributed-preprocessing.html

Each individual worker then should download the files described in the metadata files as needed, instead of assuming a shared FS. This shouldn't replace the shared FS implementation, but add S3 as an additional option.