[clang][FE][Cuda] Fix a sm90a cuda arch define check in TargetInfo (#12885)

The original commit from upstream:
[llvm-project/commit/631c6e8](llvm/llvm-project@631c6e8)
conditionally defines `__CUDA_ARCH_FEAT_SM90_ALL` separately from the
CUDA_ARCH, but we break this by making it a either or decision in a
if-else block. Hence, we were not setting correctly the definitions for
upstream's clang -x cuda execution mode for sm90a.

I believe this slipped in as a wrongly resolved merge during an upstream
pulldown.

Author: GeorgeWeb

clang/lib/Basic/Targets/NVPTX.cpp

@@ -284,10 +284,10 @@ void NVPTXTargetInfo::getTargetDefines(const LangOptions &Opts,
 
     if (Opts.SYCLIsDevice) {
       Builder.defineMacro("__SYCL_CUDA_ARCH__", CUDAArchCode);
-    } else if (GPU == CudaArch::SM_90a) {
-      Builder.defineMacro("__CUDA_ARCH_FEAT_SM90_ALL", "1");
     } else {
       Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode);
+      if (GPU == CudaArch::SM_90a)
+        Builder.defineMacro("__CUDA_ARCH_FEAT_SM90_ALL", "1");
     }
   }
 }