perf: compare native EC H2D overlap

[furionw]

Based on #14834.

Why

The vLLM CPU embedding-cache H2D overlap change needs a controlled benchmark that exposes enough cache traffic to measure concurrency with vision encoding. The prior workload used per-request random text and disabled prefix caching, so text prefill could obscure the mechanism under test. This revision makes the full text input identical, verifies that its 8K-token prefix is cached, and compares only unmodified versus patched native CPU EC on one fixed GPU pair.

Effect

Stage	Images/request	Reused after turn 1	Requests	Repetitions
Five-image	5	4	240	5
Ten-image	10	9	240	5

Both stages use concurrency 30, OSL 1024, and an 8 GiB CPU EC.

What Change

• Generate deterministic sliding-window datasets and an exact 8K-token shared prompt.
• Validate prefix-cache hits, request completion, UUID stripping, and source-tree provenance.
• Support dataset-filtered repetitions and opt-in CUDA/NVTX profiling.

classDiagram
    SweepConfig --> ServerManager : configures and launches
    ServerManager --> PrefixCacheProbe : verifies cached tokens
    SweepOrchestrator --> ServerManager : runs each arm
    SweepOrchestrator --> AIPerf : launches measured workload

    note for ServerManager "Owns server lifecycle and prefix validation"
    note for SweepOrchestrator "Sequences validation before measurement"

Loading

Test Plan

• 43 focused tests passed; five macOS-only workflow dependencies skipped.
• Run five clean repetitions per dataset, then profile both ten-image arms.