[Feature Request] Distributed Random Walks

## 🚀 Feature
We are asked for distributed (metapath-based) random walk support so I'm bookkeeping the issue here.

## Motivation
Large-scale network embedding for the following:
- [ ] DeepWalk
- [ ] metapath2vec
- [ ] node2vec (related issue: #2272 )
- [ ] (Optional) Causal Anonymous Walks that works great on dynamic graphs (in particular CTDG): http://snap.stanford.edu/caw/

Short random walks are also necessary for loss computation in negative sampling, e.g.
- [ ] GraphSAGE
- [ ] GATNE

## Alternatives

Use a third-party tool to sample the random walk paths first. May work for network embedding methods but not that straightforward for loss computation with random walks (like GraphSAGE or GATNE).

## Pitch

Efficient distributed random walk support for both homogeneous graph and heterogeneous graph (metapath-based in this case).

Start by reading this issue and the related node2vec issue (#2272), then compare the listed use cases: DeepWalk, metapath2vec, GraphSAGE, and GATNE. Done means defining and implementing efficient distributed random-walk support for homogeneous and metapath-based heterogeneous graphs, including the needs of network embedding and loss computation.

Written by the indexing model from the issue text.