Support DualPipe training

I would like to propose adding support for DualPipe in DeepSpeed. Developed by DeepSeek AI, DualPipe is a bidirectional pipeline parallelism algorithm that enhances training efficiency by fully overlapping forward and backward computation-communication phases, reducing pipeline bubbles, and optimizing resource utilization. This approach allows simultaneous processing of micro-batches from both ends of the pipeline, minimizing idle times and improving throughput, especially beneficial for training large-scale models. Integrating DualPipe into DeepSpeed would provide users with an advanced option for pipeline parallelism, complementing existing features and potentially improving training efficiency for large models. Since DualPipe is implemented with PyTorch 2.0 and above, it should integrate smoothly into DeepSpeed's existing architecture.

https://github.com/deepseek-ai/DualPipe

The issue names no repository files, tests, or entry points; begin by reading DeepSpeed's existing pipeline-parallelism implementation alongside the linked DualPipe project. Define the integration scope and validation criteria for bidirectional micro-batch processing before assessing whether support is complete.