add qwen-asr HTTP inference server

[gicrisf]

Adds a single-process HTTP inference server for Qwen3-ASR, adapted from the whisper.cpp server example.

Endpoints

• POST /inference — accepts a multipart WAV upload, returns JSON or plain text
• POST /load — hot-swap the loaded model at runtime
• GET /health — readiness probe
• GET / — serves public/index.html (live microphone UI)

Notable differences from the whisper.cpp original

• Per-request language and prompt overrides (restored to server defaults after each request)
• Response always includes inference timing and rt_factor (processing time / audio duration)
• Language dropdown in the browser UI uses Qwen3-ASR language names instead of BCP-47 codes
• No whisper-specific params (diarize, timestamps, beam search, VAD, etc.)

Build: make server → ./qwen_asr_server -d <model_dir>