Build decord from source with CUDA for GPU video decode

- Build decord with -DUSE_CUDA=ON for true NVDEC hardware decode
- Use DLPack for zero-copy transfer from decord to CuPy
- Frames stay on GPU throughout: decode -> process -> encode

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Dockerfile.gpu

@@ -26,9 +26,18 @@ RUN pip install --no-cache-dir -r requirements.txt
 # Install GPU-specific dependencies (CuPy for CUDA 12.x)
 RUN pip install --no-cache-dir cupy-cuda12x
 
-# Install PyNvCodec for true GPU-native video decoding (NVDEC)
-# Frames decode directly to GPU memory - zero CPU transfer
-RUN pip install --no-cache-dir PyNvCodec
+# Build decord from source with CUDA support for GPU-native video decoding
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    cmake build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN git clone --recursive https://github.com/dmlc/decord /tmp/decord && \
+    cd /tmp/decord && \
+    mkdir build && cd build && \
+    cmake .. -DUSE_CUDA=ON -DCMAKE_BUILD_TYPE=Release && \
+    make -j$(nproc) && \
+    cd ../python && pip install . && \
+    rm -rf /tmp/decord
 
 # Copy application
 COPY . .