[DeepSpeed] Support TP=nGPU and PP=DP=1

[comaniac]

Description

For experimental purpose, this PR supports the case that TP=#GPUs. Specifically,

1. After scheduling the model, set enable_pipeline=False so that we will use DeepSpeed engine with ZeRO-0 instead of the pipeline engine.
2. In DeepSpeed dialect, we make sure the mpu with device topology is passed to DeepSpeed engine even when pipeline is disabled, so that DeepSpeed runtime can correctly configure data parallel groups.

Checklist

• PR's title starts with a category (e.g. [Bugfix], [Model], [Tutorial], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented

cc @zarzen