Fix Windows preview mini-CTK lib layout for linker.

Merge component libs into lib/x64 or lib/arm64 so cuda_bindings can find
cudart_static.lib, and bump the prerelease cache key to drop stale flat-lib entries.

Author: Andy-Jost

.github/actions/fetch_ctk/action.yml

@@ -56,7 +56,8 @@ runs:
               "$(python "$CTK_REDIST_TOOL" preview-installer \
                 --host-platform "${{ inputs.host-platform }}" \
                 --cuda-version "${{ inputs.cuda-version }}")"
-            CTK_CACHE_COMPONENTS="${CTK_PREVIEW_INSTALLER_URL}:${CTK_PREVIEW_INSTALLER_SHA256}:${CTK_CACHE_COMPONENTS}"
+            # layout-v2: preserve lib/<arch>/ (Windows builds link against lib/x64 or lib/arm64)
+            CTK_CACHE_COMPONENTS="${CTK_PREVIEW_INSTALLER_URL}:${CTK_PREVIEW_INSTALLER_SHA256}:layout-v2:${CTK_CACHE_COMPONENTS}"
           else
             echo "CUDA prerelease packages are not supported for host-platform ${{ inputs.host-platform }}" >&2
             exit 1

ci/tools/fetch_ctk_redistrib.py

@@ -262,9 +262,11 @@ def merge_windows_preview_ctk(
         if not component_root.exists():
             raise ValueError(f"CUDA prerelease installer did not provide {component_root}")
 
+        lib_dir = destination / "lib" / arch
+
         if component == "cuda_nvrtc":
             _merge_tree(component_root / "nvrtc_dev/include", destination / "include")
-            _merge_tree(component_root / "nvrtc_dev/lib" / arch, destination / "lib")
+            _merge_tree(component_root / "nvrtc_dev/lib" / arch, lib_dir)
             _merge_tree(component_root / "nvrtc/bin" / arch, destination / "bin")
             continue
 
@@ -284,7 +286,7 @@ def merge_windows_preview_ctk(
             _merge_tree(component_root / "bin", destination / "bin")
         arch_lib = component_root / "lib" / arch
         if arch_lib.exists():
-            _merge_tree(arch_lib, destination / "lib")
+            _merge_tree(arch_lib, lib_dir)
 
     if not (destination / "include").is_dir() or not (destination / "bin/nvcc.exe").is_file():
         raise ValueError("CUDA prerelease installer did not provide the expected toolkit layout")