Fix dependency conflicts: numpy<2 override, onnxruntime cuDNN 9

- Exclude numpy and onnxruntime from CosyVoice requirements.txt
- Pin numpy<2 (onnxruntime compiled against NumPy 1.x ABI)
- Use onnxruntime-gpu>=1.19 (cuDNN 9 compatible, base image has cuDNN 9)
- Remove redundant numpy from API dependencies line

Author: hsiang-han

docker/gpu/Dockerfile

@@ -25,14 +25,14 @@ RUN pip install --no-cache-dir transformers==4.51.3
 # using the system setuptools which still has pkg_resources
 RUN pip install --no-cache-dir "setuptools<71" wheel cython "numpy<2" && \
     cd cosyvoice_repo && \
-    grep -iv "torch\|torchaudio\|tensorrt\|deepspeed\|transformers\|openai-whisper" requirements.txt > requirements_filtered.txt && \
+    grep -iv "torch\|torchaudio\|tensorrt\|deepspeed\|transformers\|openai-whisper\|onnxruntime\|numpy" requirements.txt > requirements_filtered.txt && \
     pip install --no-cache-dir -r requirements_filtered.txt && \
     pip install --no-cache-dir --no-build-isolation openai-whisper==20231117 && \
-    pip install --no-cache-dir modelscope onnxruntime-gpu huggingface_hub && \
+    pip install --no-cache-dir modelscope "onnxruntime-gpu>=1.19" huggingface_hub && \
     pip install --no-cache-dir --force-reinstall torchvision --index-url https://download.pytorch.org/whl/cu128
 
-# Install API dependencies
-RUN pip install --no-cache-dir fastapi uvicorn[standard] python-multipart numpy soundfile
+# Install API dependencies (numpy already pinned <2 above, don't override)
+RUN pip install --no-cache-dir fastapi uvicorn[standard] python-multipart soundfile
 
 # Copy API code
 COPY api/ ./api/