feat(qwen3-asr): add offline vLLM example

[LauraGPT]

Summary

• add an offline Qwen3-ASR vLLM example based on the native Qwen3ASRModel.LLM backend
• normalize arbitrary input audio to mono 16 kHz, use qwen-asr native chunking, and retain exact chunk timestamps in JSON output
• add Chinese and English operational notes, including the distinction from FunASR AutoModelVLLM and accuracy-comparison boundaries
• add focused tests for timestamps, result-count validation, lazy optional dependencies, and CLI defaults

Related to #3419. This PR intentionally does not close the issue because the reporter still needs to validate the workflow in their environment.

Verification

• 9 passed: the new tests plus existing Qwen3-ASR WebSocket and dependency-check tests
• py_compile, CLI --help, and git diff --check passed
• real NVIDIA H100 80 GB run with Qwen/Qwen3-ASR-1.7B on the 480.24 s 離線轉錄的準確率不佳 #3419 attachment produced three monotonic chunks covering 0-480.216 s
• output evidence SHA256: 899d0eacc8db73e22cbe7e7aeb8befd56c8cc775898862aa327dc4d90a7df7aa

Accuracy note

The earlier 18.74% prefix-trim CER used the original reference and a standard character normalization. The reporter later supplied an updated reference containing the first 17.69 seconds and their exact scorer (SHA256 66ec56383c505418c78c5e7fe567d33564fcce7faccb0f508959a57f76d4ae50), which additionally normalizes numbers and ignores selected one-character fillers and substitutions. Under that exact custom scoring contract, the PR's native vLLM output scores 6.45% CER; the separately tested Transformers pipeline scores 6.80%, and the 20-second VAD/vLLM path scores 6.63%. These results reproduce the reporter's stated 8-9% range but are not directly comparable with the earlier standard-normalization result or an official benchmark.

Attribution

The architecture was cross-checked against the MIT-licensed LanceLRQ/qwen3-asr-service project, which is acknowledged in the accompanying notes.