Measure H100 PCIe and H100 SXM

[parda21]

We do not have an H100. Projection (Ampere issue rates, docs/results.md section 3): PCIe ~1.1-1.2x, SXM ~0.8-0.9x, which is why the tensor-core path exists as a plan. Needed: python -m nwc.demo Parda21/Qwen3-4B-NWC --load --graph, the same with Qwen/Qwen3-4B --native, and from a clone scripts/kernbench.py --runs 30 plus experiments/pipe_bench.cu for the Hopper instruction costs. Post results in a benchmark issue or here.