docs: add tested native Transformers pipeline entry point

[LauraGPT]

Summary

Document the actual registered any-to-any pipeline entry point for native Fun-ASR-Nano. Include a pinned public audio CPU example, explicit generation config and the checkpoint's structured language field. Explain why the standard automatic-speech-recognition route is not interchangeable on Transformers 5.17.0.

Evidence

• Unpatched stable5.17.0 + torch2.10.0+cpu: standard ASR pipeline fails at embedding because float audio features reach token indices; registered any-to-any transcribes the same pinned English recording.
• Exact README Python fence executed against the pinned public audio URL and official cached weights: expected English text returned. No remote code or token required.
• One new AST contract test RED before the snippet, full32 tests pass after and on this exact signed commit.
• Independent read-only review of both final hashes found no P1/P2.
• CPU CLI/GPU guide remain unchanged. No pipeline GPU/batch/quality/capacity claim. Attention warning remains observed.

Source bundle, refs, raw failing/successful probes and exact source hashes backed up on ind-gpu8. Signed commit and DCO. No runtime/model/package change.