[build] Only bundle CUDA DLLs that cv2.pyd actually needs

PE import analysis shows ~1.3 GB of CUDA toolkit DLLs bundled in the
wheel are never referenced by cv2.pyd or its transitive dependencies.
Switch from copying all CUDA bin/*.dll to an allowlist of only the
DLLs that are actually linked: cublas, cublasLt, cudart, cufft, and
the npp libraries.

Author: Breakthrough

.github/workflows/build_wheels_windows.yml

@@ -124,7 +124,16 @@ jobs:
         echo "Adding CUDA to PATH..."
         $CUDA_PATH = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\${{ matrix.cuda-path-version }}"
         echo "CUDA_PATH=$CUDA_PATH" | Out-File -FilePath $env:GITHUB_ENV -Append
-        Copy-Item -Path "$CUDA_PATH/bin/*" -Destination .  -Include "*.dll"
+        # Only copy CUDA DLLs that cv2.pyd actually needs (determined via PE import analysis).
+        # This excludes ~1.3 GB of unused libraries (cusparse, cusolver, curand, nvrtc, etc.)
+        $needed_dlls = @(
+          "cublas64_*", "cublasLt64_*", "cudart64_*", "cufft64_*",
+          "nppc64_*", "nppial64_*", "nppicc64_*", "nppidei64_*", "nppif64_*",
+          "nppig64_*", "nppim64_*", "nppist64_*", "nppitc64_*"
+        )
+        foreach ($pattern in $needed_dlls) {
+          Copy-Item -Path "$CUDA_PATH\bin\$pattern.dll" -Destination .
+        }
       shell: pwsh
     - name: 🔧 Install NVIDIA CuDNN
       run: |