fix: make clang recognize nvvm repo as valid cuda installation

[cloudhan]

Together with #431, this should fix hermetic ctk13 with clang as cuda compiler driver.

cd tests/integration/toolchain_redist_json
sed -i 's/12\.6\.3/13\.0\.0/g' MODULE.bazel
export USE_BAZEL_VERSION=8.5.0
export CUDA_COMPILER_USE_CC_TOOLCHAIN=true
bazel build --repo_env=CC=clang --@rules_cuda//cuda:compiler=clang --@rules_cuda//cuda:archs=sm_89 :use_rule

produces

ERROR: /home/cloud/rules_cuda/tests/integration/toolchain_redist_json/BUILD.bazel:10:13: Compiling external/rules_cuda_examples+/basic/kernel.cu failed: (Exit 1): clang failed: error executing CudaCompile command (from target //:use_rule) /usr/lib/llvm-22/bin/clang -x cu '--cuda-gpu-arch=sm_89' '--cuda-path=external/rules_cuda++toolchain+cuda_nvcc_v13.0.48/nvcc' '-frandom-seed=bazel-out/k8-fastbuild/bin/_objs/use_rule/kernel.pic.o' ... (remaining 76 arguments skipped)

Use --sandbox_debug to see verbose messages from the sandbox and retain the sandbox build root for debugging
clang: error: cannot find libdevice for sm_89; provide path to different CUDA installation via '--cuda-path', or pass '-nocudalib' to build without linking with libdevice
clang: error: cannot find CUDA installation; provide its path via '--cuda-path', or pass '-nocudainc' to build without CUDA includes
clang: error: cannot find CUDA installation; provide its path via '--cuda-path', or pass '-nocudainc' to build without CUDA includes
Target //:use_rule failed to build

The reason behind this is https://github.com/llvm/llvm-project/blob/6de6f7b46bda0fef45b997d05fe57e046f60e4df/clang/lib/Driver/ToolChains/Cuda.cpp#L206-L290

The <ctk_root> must have the structure as follows,

<ctk_root>/bin/
<ctk_root>/include/
<ctk_root>/nvvm/libdevice/libdevice.10.bc

then <ctk_root> can be treat as a valid CTK installation.

We fake it by patching nvvm repo which contains libdevice.10.bc with empty bin/ and include/