[pytorch] expose __ldg(const Half* ptr) to Clang in host mode (pytorch#38151)

Summary:
Pull Request resolved: pytorch#38151

We need to expose this method to Clang unconditionally when building CUDA, otherwise it would error on device code calling `__ldg` with `Half*`.

Test Plan:
```
buck build -c fbcode.caffe2_use_mpi=1 -c fbcode.cuda_use_clang=true mode/opt //experimental/training_supercomputer/trainer/hpc_pt:trainer
```

Reviewed By: ngimel

Differential Revision: D21481297

fbshipit-source-id: aacfe7de2cdc8542908249081ddb58170b1e35ff

Author: Igor Sugak

c10/util/Half-inl.h

@@ -45,7 +45,8 @@ inline C10_HOST_DEVICE Half::operator __half() const {
 
 // CUDA intrinsics
 
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 350)
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 350)) || \
+    (defined(__clang__) && defined(__CUDA__))
 inline __device__ Half __ldg(const Half* ptr) {
     return __ldg(reinterpret_cast<const __half*>(ptr));
 }