[vLLM-ATOM] Add profile trace parsing tool for vLLM-ATOM

[kliuae-amd]

Motivation

This PR adds a profile trace parsing tool that breaks down vLLM-ATOM traces for hot kernel analysis.

Technical Details

Provide kernel breakdowns of mixed/prefill and decode batches for kernel runtime analysis.
It leverages vLLM's trace markers enabled by VLLM_CUSTOM_SCOPES_FOR_PROFILING=1 to gather kernels in batches.
Currently it uses input layernorms as boundaries to separate layers from one another.

Usage:

python parse_vllm_trace.py <trace.json.gz>
    [--layer N]            Target transformer layer index.
    [--percentile P]       The X-percentile batches to extract kernel stats from.
    [--eager-trace <path>] Companion eager trace for extracting host module attributes.
                           For graph-captured batches like decode workloads in FULL_AND_PIECEWISE settings,
                           GPU kernels in a graph are packed into a single graph replay event.
                           An eager trace can be provided alongside with graphed traces to establish kernel
                           launches with host side launch modules from eager traces.
    [--output-prefix <str> Prefix for output XLSX filename.

Test Plan

Test Result

Submission Checklist

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