docs(vllm): add measured L20 tuning profile

[LauraGPT]

Summary

• document the L20 / 47-second / 16-client tuning point reported and profiled in 1.4.3 的 funasr/bin/realtime_ws.py 的 run_session_work() 实现太简单粗暴，导致并发性能比1.3.9还差很多 #3528
• provide a copyable partial-window-sec=8, decode-interval=2.0 command with decode profiling
• state the exact workload and measured latency/throughput boundaries
• keep the result scenario-specific instead of changing or implying a new global default
• update English and Chinese production guides together

Verification

• 21 documentation and realtime benchmark tests passed
• bilingual content contract passed
• git diff --check passed

Addresses the documentation follow-up in #3528; the issue should remain open for adaptive refresh work and broader hardware evidence.

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com