[CUDA] Fix Protobuf warning in CUDA 13.3 plugin build (#32854)
## Description
CUDA 13.3 reports diagnostic 68 for Protobuf's intentional conversion of
a signed enum pointer mask to `uintptr_t`. The CUDA plugin build treats
all NVCC warnings as errors, so this third-party header diagnostic
prevents the plugin LLM object library from compiling.
## Summary of Changes
### CUDA plugin build
| File | Change |
|------|--------|
| `cmake/onnxruntime_providers_cuda_plugin.cmake` | Suppress NVCC
diagnostic 68 for CUDA 13+ plugin targets and document the Protobuf
construct that triggers it. |
The suppression is limited to CUDA compilation and follows the existing
CUDA 13 workarounds for diagnostics emitted by Protobuf and Abseil
headers.
## Testing
- `D:\py312\Scripts\lintrunner.exe -a
cmake/onnxruntime_providers_cuda_plugin.cmake`
- `D:\py312\Scripts\cmake.exe --build build\cuda13_plugin\Release
--config Release --target onnxruntime_providers_cuda_plugin_llm --
/maxcpucount:1 /p:CL_MPCount=1 /nodeReuse:False`
- `.build\cuda13_plugin.cmd` using CUDA 13.3; completed successfully and
produced `onnxruntime_gpu-1.30.0-cp312-cp312-win_amd64.whl`
## Checklist
- [x] Build validation completed
- [x] No breaking changes
- [x] No documentation changes required