docs: add NVFP4 conversion and benchmarking notes

[MerkyorLynn]

Adds a short doc page (docs/nvfp4.md) on the NVFP4 weight format: how to obtain an NVFP4 GGUF, when the native FP4 path is selected, and how to benchmark it against Q4_K_M.

NVFP4 type/CUDA support landed in #19769 (ggml type) and #22196 (Blackwell native NVFP4), but there's no doc covering the practical questions: where NVFP4 GGUFs come from (converting ModelOpt / compressed-tensors checkpoints via convert_hf_to_gguf.py), when the native FP4 path activates (blackwell_mma_available()), and NVFP4 vs Q4_K_M for prompt processing vs single-stream decode.

Avoids absolute performance/quality claims — points to llama-bench and task-specific eval, since results depend on model, build, and workload. Also adds a one-line pointer from tools/quantize/README.md. No code changes.

[ggml-gh-bot]

Hi @MerkyorLynn, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[MerkyorLynn]

Thanks for the reminder. AI assistance was used to help draft and revise this documentation, but I manually reviewed the content against the current llama.cpp code paths and PR history before submitting. I can explain and revise any part of the doc if maintainers have concerns.

[MerkyorLynn]

Closing this for now so I only keep one open PR as a new contributor. I may revisit the NVFP4 docs later after the current MTP tuning PR is resolved.

[MerkyorLynn]

Closing this for now so I only keep one open PR as a new contributor. I may revisit the NVFP4 docs later after the current MTP tuning PR is resolved.