Integrate fast.cu kernels and B200 optimization experiments

[cfregly]

Add all 34 fast.cu source files and four benchmark pairs: H100 BF16 GEMM, H100/B200 int32 reduction, B200 epilogue stores, and GB300 NVFP4. The guide maps the examples to Chapters 9, 10, 11, and 14. Sources are pinned to 2dfe5e26aecfd9e5f27bf9d5837deea01acda24b with the MIT license.

The adapters use the caller's CUDA stream and random seed, check every output, and keep compilation and preparation outside timing. The reduction adapter fixes the output-reset race. The B200 epilogue compares 128-bit and 256-bit stores with the same cache policy.

Validation:

• B200 tests: 66 passed and 2 hardware skips. Local tests, including README generation: 75 passed and 5 hardware skips.
• Benchmark checks covered 944 entries with no errors or warnings. Source hashes, links, formatting, syntax, and YAML checks passed.
• Both B200 pairs passed input/output checks and collected Nsight Systems and Compute profiles. One exploratory VM run measured 1.0649x for the epilogue and 1.0455x for reduction. Reduction missed the required 1.05x threshold. These results were not added as expected speedups.
• The H100 adapter compiled and linked for SM90a. NVFP4 r0-r9 compiled, linked, and imported together for SM103 with CUDA 13.1.80. No H100 or SM103 GPU kernels ran.

The NVFP4 work fixes adapter build and import issues. It does not change the newer upstream kernels or port their full GEMM to B200. Details are in code/labs/fast_cu/validation.md.

The dashboard dependency audit reported existing advisories in Next.js, sharp, and js-yaml. Its package files matched main and were unchanged by this PR.