Switch to CUDA 12.8 + ONNX-based VAD for RTX 5090 Blackwell support

Upgrade PyTorch to 2.7+ with cu128 wheels for Blackwell (sm_120) GPU
support. Replace silero-vad (which depends on torchaudio) with a direct
ONNX Runtime implementation of the same Silero VAD model, eliminating
the torchaudio dependency entirely.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

server/models.py

@@ -41,11 +41,15 @@ class ModelManager:
         log.info("All models loaded successfully.")
 
     def _load_vad(self):
-        log.info("Loading Silero VAD...")
-        from silero_vad import load_silero_vad
+        log.info("Loading Silero VAD (ONNX)...")
+        from huggingface_hub import hf_hub_download
+        from server.vad import SileroVADOnnx
 
-        self.vad_model = load_silero_vad()
-        log.info("Silero VAD loaded (CPU).")
+        model_path = hf_hub_download(
+            repo_id="onnx-community/silero-vad", filename="silero_vad.onnx"
+        )
+        self.vad_model = SileroVADOnnx(model_path)
+        log.info("Silero VAD loaded (ONNX, CPU).")
 
     def _load_asr(self):
         log.info("Loading Qwen3-ASR-0.6B (transformers backend)...")