refactor: replace Kokoro-82M with suno/bark for TTS, update HF cache path, and add model warmup on startup.

test_cross.py

@@ -0,0 +1,20 @@
+import asyncio
+from backend.tts_asr import _tts_sync_with_retry, _load_asr_pipeline_with_retry, _asr_pipeline
+import base64
+
+async def main():
+    text = "早上好"
+    print(f"Testing TTS with text: {text}")
+    audio_bytes, sr = await _tts_sync_with_retry(text, rate=1.0)
+    print(f"TTS generated {len(audio_bytes)} bytes of audio.")
+    
+    print("Testing ASR...")
+    await _load_asr_pipeline_with_retry()
+    asr = _asr_pipeline
+    
+    # Needs to process audio_bytes. ASR expects float32 numpy array or bytes?
+    # the pipeline takes bytes or dict with raw array
+    result = asr(audio_bytes) 
+    print("ASR output:", result)
+
+asyncio.run(main())