docs(vllm): clarify official model paths

[LauraGPT]

Summary

• clarify that the official FunAudioLLM/Fun-ASR-Nano-2512 checkpoint stores the complete model.pt at the repository root
• explain that Qwen3-0.6B/ is a config/tokenizer directory and that prepare_vllm_model_dir() creates Qwen3-0.6B-vllm/model.safetensors
• distinguish the FunASR split-engine path from the separate native vLLM FunASRForConditionalGeneration community checkpoint
• keep the English and both Chinese vLLM guides aligned with a regression contract test

Why

The previous guide did not explain why the official checkpoint Qwen3-0.6B/ directory has no weights, which made it look incomplete and encouraged users to mix two incompatible model paths.

Validation

• python3 -m pytest tests/test_vllm_model_source_docs.py tests/test_docs_funasr_install_commands.py tests/test_canonical_qwenaudio_links.py -q: 21 passed
• Black pre-commit hook: passed
• git diff --check: passed
• official Hugging Face revision 272c57b82523ada6fd87095e955f8e29100979ab exercised on an H100 with vLLM 0.19.1
• automatic extraction produced a 1,503,300,296-byte model.safetensors
• real example/zh.mp3 transcription: 开饭时间早上九点至下午五点。

Closes #3475