Add PyTorch/ONNX prediction models with physics fallback

- Train 3 MLP networks (acid speed 14→1, tension 4→10, quality 6→2)
  on 12,000 synthetic samples generated from physics models + noise
- Export pre-trained ONNX models to pt_models/ directory
- Rewrite prediction.py: ONNX inference first, physics fallback if unavailable
- Add onnxruntime + numpy to requirements.txt (Aliyun mirror for Docker)
- Use Tsinghua mirror in Dockerfile for pip installs

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

backend/requirements.txt

@@ -18,3 +18,5 @@ redis==5.0.4
 aioredis==2.0.1
 httpx==0.27.0
 loguru==0.7.2
+onnxruntime==1.18.0
+numpy==1.26.4