Fix CUDA CI: detect pre-installed toolkit, fallback to cuda-toolkit-12-6

The Jimver/cuda-toolkit action failed because cuda-curand-dev-12-6
package doesn't exist. The GPU runner image (NVIDIA GPU-Optimized
Image for AI and HPC) likely has CUDA pre-installed. Now we:
1. Check if nvcc is already on PATH
2. Check /usr/local/cuda and add to PATH
3. Fallback to installing cuda-toolkit-12-6 meta-package

Author: lukekim

.github/workflows/ci_cuda.yaml

@@ -23,12 +23,21 @@ jobs:
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
-      - name: Install CUDA toolkit
-        uses: Jimver/cuda-toolkit@v0.2.21
-        with:
-          cuda: '12.6.3'
-          method: 'network'
-          sub-packages: '["nvcc", "cudart-dev", "cublas-dev", "curand-dev"]'
+      - name: Setup CUDA toolkit
+        run: |
+          # The GPU runner image has CUDA pre-installed, just ensure PATH is set
+          if command -v nvcc &>/dev/null; then
+            echo "CUDA already available: $(nvcc --version | grep release)"
+          elif [ -d /usr/local/cuda ]; then
+            echo "/usr/local/cuda/bin" >> $GITHUB_PATH
+            echo "CUDA found at /usr/local/cuda"
+          else
+            echo "Installing CUDA toolkit..."
+            sudo apt-get update
+            sudo apt-get install -y cuda-toolkit-12-6
+            echo "/usr/local/cuda-12.6/bin" >> $GITHUB_PATH
+            echo "LD_LIBRARY_PATH=/usr/local/cuda-12.6/lib64:$LD_LIBRARY_PATH" >> $GITHUB_ENV
+          fi
       - name: Install dependencies
         run: sudo apt-get update && sudo apt-get install -y curl build-essential libssl-dev protobuf-compiler pkg-config
       - name: Install Rust Stable