Add ReFlow ZeRO-3 CPU-offload fine-tuning examples

[st-bang97]

Companion examples for deepspeedai/DeepSpeed#8719, which implements
deepspeedai/DeepSpeed#8697.

Summary

This PR adds training/DeepSpeed-Reflow, a set of end-to-end fine-tuning
examples for ReFlow, the asynchronous CPU-offloaded optimizer path
for DeepSpeed ZeRO-3.

The examples cover:

• OPT and Llama fine-tuning on single- and multi-GPU configurations.
• Adam/AdamW and Lion optimizer paths.
• ReFlow and the corresponding plain ZeRO-3 CPU optimizer-offload
  baseline using otherwise matching configurations.
• ReFlow configuration through the zero_optimization.reflow block.
• Throughput and memory comparison between ReFlow and the baseline.
• Deterministic loss-equivalence checking.
• Launcher, configuration, failure-propagation, and preprocessing tests.

The runtime implementation itself is provided by
deepspeedai/DeepSpeed#8719. This PR contains only the example,
benchmark, validation, and usage code.

Examples

The added launchers include:

• OPT-350M on 1 GPU
• OPT-30B on 1, 4, and 8 GPUs
• Llama-7B, Llama-8B, and Llama-13B on 1 GPU
• Llama-70B on 8 GPUs
• Adam/AdamW and Lion variants for the larger multi-GPU examples

Each launcher accepts either reflow or zerooffload mode, allowing
the same model and training configuration to be run with and without
ReFlow.

For example:

cd training/DeepSpeed-Reflow

bash finetune_opt-30b_8gpu.sh reflow
bash finetune_opt-30b_8gpu.sh zerooffload