docs(examples): add GB10/DGX Spark NVFP4 optimized sweep specs

[dutifulbob]

Opened on behalf of Onur Solmaz (osolmaz).

Summary

We tuned how vLLM is launched for four NVFP4 models on a GB10 (DGX Spark) and got noticeably higher throughput, but that config only lived in local run dirs.
This adds those tuned sweep specs to the repo as a versioned example, plus a README explaining what changed and why, and a record of the old vs new serve command per model.
It matters because the original hand-authored specs launched vLLM in a way that quietly hurt decode, and now the better config is reproducible instead of lost.

What Changed

Adds a new example directory with the four tuned specs and documentation. No library code changes.

• examples/gb10-dgx-spark-nvfp4-optimized/ with qwen27b.json, qwen35b.json, gemma4-26b.json, gemma4-31b.json.
• README.md explaining the tuning: drop --mamba-cache-mode none (the Qwen models are hybrid linear-attention and that flag disabled caching the recurrent state), drop --enforce-eager (enables CUDA graphs), drop --no-async-scheduling, and raise --max-num-batched-tokens to 16384.
• PROFILES-old-vs-new.txt recording each model's old serve command vs the tuned one.
• The README also notes two tested-but-not-adopted levers: MTP (helps single-stream decode ~1.6x but hurts aggregate, so off) and flashinfer autotune (no throughput gain on sm_121).

Testing

Confirmed the specs are valid and the docs pass the repo's checks. Behavior was measured on the real machine before committing.

• All four specs round-trip localperf bench plan (39/43/39/39 planned runs).
• npx @simpledoc/simpledoc check — clean.
• go build ./... — ok (no code changed).
• Live: ran the tuned sweeps end to end; qwen27b aggregate output rose ~20-30% at concurrency (4k-ref c32 157 -> 186 tok/s, decode-8k c32 61 -> 80, decode-16k c16 31 -> 40).

Risks

Low — this is example data and docs, no code paths change.

• The specs are machine-specific: runner.vllm_command, env (CUTE_DSL_ARCH=sm_121a), and ports point at one host. The README says to edit before use elsewhere, matching the existing example convention.

Follow-ups

• Optional: a standalone docs/ tuning note if we want the findings outside the example dir.