Native Windows CUDA build rejects unguarded GCC diagnostic pragmas

The first native Windows SM86 build of SPEC-DFLASH2 W3 passes its repaired CUDA translation units, then fails while compiling src/vllm/platforms/cuda.cpp with MSVC /WX.

Observed errors:

src/vllm/platforms/cuda.cpp(231): warning C4068: unknown pragma 'GCC'
src/vllm/platforms/cuda.cpp(232): warning C4068: unknown pragma 'GCC'
src/vllm/platforms/cuda.cpp(243): warning C4068: unknown pragma 'GCC'
error C2220: warning treated as error

Commit cea682956 added unguarded #pragma GCC diagnostic directives around the CUDA platform registrar. This repository treats warnings as errors and supports native Windows CUDA, so the GCC-only pragmas require a compiler guard. A separate Qwen3.8 development branch already carries the expected #if defined(__GNUC__) shape, but that branch is not main and cannot satisfy this row's immutable build.

Owner: SPEC-DFLASH2 W3 repair flow. Add a focused compile/contract regression, preserve the GCC suppression, and run the Windows CUDA focused build through link before closing.

Start in src/vllm/platforms/cuda.cpp around the CUDA platform registrar and compare the GCC diagnostic pragma shape with commit cea682956 and the referenced Qwen3.8 development branch. Add the focused compile/contract regression requested in the issue, then run the native Windows CUDA focused build through linking and confirm the warning-as-error failure is gone.