docs(nano): clarify training-speed evidence and release scope

[LauraGPT]

Summary

Clarify the Fun-ASR-Nano fine-tuning guide after #3705 merged. Related issue: #3704 (not closed by this docs-only change).

• State that the two training-speed options are absent from the published funasr==1.4.16 wheel and require matching source package/recipe revisions.
• Attribute cuDNN planning observations to the reported single-B200/PyTorch 2.11.0/cuDNN 9.19 setup, rather than guaranteeing behavior across sm_90/sm_100 workloads.
• Preserve the process-wide SDPA warning; explain compilation startup cost, the contributor's 23/6,128 out-of-tolerance checks, untested multi-GPU/DeepSpeed scope, and how to disable/isolate the options.

Only the existing Training-speed options section changes. No runtime code, recipe defaults, model config, dependency, release or deployment changes.

Evidence

• Verified current source and the actual PyPI 1.4.16 wheel against published SHA-256 f95943f6a8111349a1b08907363b048a49e9cb9caa1e3afd8c92970d08bf450d; the wheel lacks funasr/models/fun_asr_nano/llm_forward_opts.py and the model wiring.
• Checked claims against the merged [perf] Fun-ASR-Nano fine-tuning on B200: 0.244 s to 0.068 s per step at 1 GPU, mostly from cuDNN SDPA plan builds on every new batch shape #3705 body, its source helper and recipe.
• Targeted guide checks passed: unchanged content outside this section, preserved option list/shared-process warning, relative links, source/wheel availability boundary, and GitHub GFM rendering.
• python -m pytest -q tests/test_training_docs_contract.py -k 'not no_download_toy': 5 passed, 1 deselected. The toy-model execution case was outside this docs-only validation; no full-suite or training acceptance is claimed.
• git diff --check passed; signed commit and DCO sign-off verified locally and on GitHub.
• Independent source/document review found no P1/P2 issues at document SHA-256 8d192f0b68c2462cd30118524b229535ea628932a293f133292be7f8166a5b19.

No new B200/H100 benchmark, model inference, numerical-equivalence result, or universal speedup is claimed. The numerical/performance observations remain the contributor's explicitly scoped results.

AI-assisted: prepared and reviewed with Codex; source, published wheel contents, rendered guide, and scoped checks were verified before submission.

CI Scope

No workflow is triggered by this single example-document path under the current workflow filters. Exact-head checks, commit statuses, and Actions runs are empty; this is not a claim that CI passed. The local checks and GFM rendering above provide the executed validation. No workflow, required check, or branch protection was changed.