added vllm_disagg profiling

[AakarshAMD]

Motivation

Add an opt-in profiling workflow for vLLM disaggregated serving using rocprofv3, ROCtx, and MoRIIO instrumentation while preserving the normal non-profiling workflow.

Technical Details

• Added a dedicated ROCm 7.2.3 profiling Dockerfile with rocprofv3 and MoRIIO request-ID instrumentation.
• Added benchmark_xPyD_profile.sh for profiling-specific request IDs, timing manifests, warmup control, and bounded workloads.
• Added conditional kernel/marker capture and graceful worker finalization.
• Added composite request-ID correlation, combined traces, request statistics, and kernel categorization.
• Added true local-rank-0 selection and cross-node clock alignment for combined traces.
• Added validation for capture completeness, malformed durations, request identity, MoRI transfer attribution, and analysis failures.
• Profiling requires RUN_PROFILE=1 with CONNECTOR=moriio; normal runs remain unchanged.

Test Plan

• Build the dedicated profiling image using the profiling target.
• Run profiled 1P1D and 2P2D configurations across TP and EP models.
• Verify benchmark completion, capture finalization, request-ID consistency, ROCtx markers, MoRI attribution, combined traces, and kernel-analysis outputs.

Test Result

All validation jobs below passed on their first submitted attempt:

• Llama 3.3 70B FP8-KV — 1P1D, TP8 (jobs 209995 and 210278)
• GPT-OSS 120B — 1P1D, TP8 (job 210279)
• DeepSeek-R1 — 1P1D, EP8 (job 210280)
• DeepSeek-V3 — 1P1D, EP8 (job 210281)
• DeepSeek-R1 — 2P2D, EP16/stage (job 210282)
• DeepSeek-V3 — 2P2D, EP16/stage (job 210283)

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.