perf(opsd): benchmark shared prefill with CUDA Graph decode - #1010

Summary

Extend the OPSD HybridEngine benchmark with:

• Prefill/decode forward timing
• Native CUDA Graph decode benchmarking
• Combined shared-prefill and CUDA Graph benchmarking
• JSON mode metadata and documentation

This accompanies deepspeedai/DeepSpeed#8350 and the shared-prefill CUDA Graph follow-up.

Validation

• Shared-prefill CUDA Graph smoke benchmark: passed
• Four-mode OPT-6.7B benchmark completed: eager, shared, graph, shared + graph
• Combined result: 10.04% lower total latency and 11.16% higher throughput
• Commit is signed off

Thank you for reviewing!