Fix CUDA backend Windows build under C++20 (pytorch#23135)

### Summary
All \`test-model-cuda-windows-e2e\` jobs on main fail to build:

\`\`\`
backends\cuda\runtime\platform\platform.cpp(63,20): error C2440:
'initializing': cannot convert from 'std::basic_string<char8_t,...>' to
'std::basic_string<char,...>'
\`\`\`

pytorch#22793 made \`executorch_core\` export \`cxx_std_20\`, so
\`cuda_platform\` now compiles as C++20, where
\`std::filesystem::path::u8string()\` returns \`std::u8string\`, which
no longer converts to \`std::string\`.

Fix: load the library with \`LoadLibraryW(path.c_str())\` (which is also
correct for non-ASCII paths), and build a UTF-8 \`std::string\` only for
the error log. The conversion compiles under both C++17 and C++20.


**Second fix:** once platform.cpp built, `aoti_cuda_shims` failed next:

```
c10/util/safe_numerics.h(71): error : namespace "std" has no member "countl_zero"
```

The `.cu` shims are compiled by nvcc, and `backends/cuda/CMakeLists.txt`
pinned `CMAKE_CUDA_STANDARD 17`. `cxx_std_20` on `executorch_core` only
raises the CXX language, so it never reaches nvcc. The CUDA backend now
uses C++20 for both CXX and CUDA on Windows (matching the top-level
rule) and stays on C++17 elsewhere.

### Test plan
CI: \`Test CUDA Windows Export and E2E\` (triggered by the
\`backends/cuda/**\` path filter).

Author: Gasoonjia

backends/cuda/CMakeLists.txt

@@ -16,9 +16,17 @@
 #
 cmake_minimum_required(VERSION 3.26)
 
-set(CMAKE_CXX_STANDARD 17)
+# The .cu shims include vendored c10's safe_numerics.h, whose MSVC-only branch
+# needs C++20 (see the top-level CMakeLists.txt). cxx_std_20 on executorch_core
+# does not reach nvcc, so the CUDA standard has to be raised here too.
+if(WIN32)
+  set(CMAKE_CXX_STANDARD 20)
+  set(CMAKE_CUDA_STANDARD 20)
+else()
+  set(CMAKE_CXX_STANDARD 17)
+  set(CMAKE_CUDA_STANDARD 17)
+endif()
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CUDA_STANDARD 17)
 set(CMAKE_CUDA_STANDARD_REQUIRED ON)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)

backends/cuda/runtime/platform/platform.cpp

@@ -60,10 +60,12 @@ namespace cuda {
 executorch::runtime::Result<void*> load_library(
     const std::filesystem::path& path) {
 #ifdef _WIN32
-  std::string utf8 = path.u8string();
-  auto lib_handle = LoadLibrary(utf8.c_str());
+  auto lib_handle = LoadLibraryW(path.c_str());
   if (lib_handle == NULL) {
     const DWORD err = GetLastError();
+    // u8string() returns std::u8string under C++20, so copy into std::string.
+    const auto u8 = path.u8string();
+    const std::string utf8(u8.begin(), u8.end());
     ET_LOG(
         Error,
         "Failed to load %s with error %lu: %s",