Add OpenAI-compatible ASR API server

Add asr-server binary with axum-based HTTP server exposing:
- POST /v1/audio/transcriptions (multipart: file, language, response_format)
- GET /v1/models
- GET /health

Supports json, text, and verbose_json response formats. CLI args:
--model-dir, --host, --port, --language, -v. Uses spawn_blocking
with Arc<Mutex<AsrInference>> for thread-safe inference dispatch.

Also adds duration_seconds to TranscribeResult and includes asr-server
in release CI packaging.

Signed-off-by: Michael Yuan <michael@secondstate.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: juntao

.github/workflows/release.yml

@@ -155,6 +155,7 @@ jobs:
         run: |
           mkdir ${{ matrix.asset-name }}
           cp target/release/asr ${{ matrix.asset-name }}/asr
+          cp target/release/asr-server ${{ matrix.asset-name }}/asr-server
           cp -r tokenizers ${{ matrix.asset-name }}/tokenizers
           if [ "${{ matrix.backend }}" = "mlx" ]; then
             cp target/release/build/qwen3_asr-*/out/lib/mlx.metallib ${{ matrix.asset-name }}/