Make a more general between-stage communication class/abstraction

[zhisbug]

In our current design of 3D parallelism, there are possibilities that:

• stage 0 working with a mesh with 1 node 2 GPUs
• stage 1 working with a mesh with 2 node 2 GPUs

on each mesh, the model might be sharded differently.

The communication between stage0 and stage1 is not necessarily based on device-device send/recv, because it might involve communications from multiple devices to multiple devices.

We should make some abstraction in the pipeline runtime to handle this communication pattern

[zhisbug]

This issue will tackle #25

[zhisbug]

#47