Add C10_EMBEDDED to gate ostream usage in Half/BFloat16 (pytorch#140566)

We want to use Half/BFloat16 in ExecuTorch to support shared kernel code. They will need to be used in ExecuTorch core, so they can't have streams. This diff introduces a macro to gate the stream code off.

Differential Revision: [D65888035](https://our.internmc.facebook.com/intern/diff/D65888035/)
Pull Request resolved: pytorch#140566
Approved by: https://github.com/ezyang, https://github.com/malfet
ghstack dependencies: pytorch#140564, pytorch#140565

Author: swolchok

c10/util/BFloat16.h

@@ -8,7 +8,9 @@
 #include <cstdint>
 #include <cstring>
 #include <iosfwd>
+#ifndef C10_EMBEDDED
 #include <ostream>
+#endif // C10_EMBEDDED
 
 #if defined(__CUDACC__) && !defined(USE_ROCM)
 #include <cuda_bf16.h>
@@ -114,12 +116,14 @@ struct alignas(2) BFloat16 {
 #endif
 };
 
+#ifndef C10_EMBEDDED
 C10_API inline std::ostream& operator<<(
     std::ostream& out,
     const BFloat16& value) {
   out << (float)value;
   return out;
 }
+#endif // C10_EMBEDDED
 
 } // namespace c10
 

c10/util/Half.h

@@ -29,7 +29,9 @@
 #include <cstring>
 #include <iosfwd>
 #include <limits>
+#ifndef C10_EMBEDDED
 #include <ostream>
+#endif // C10_EMBEDDED
 
 #ifdef __CUDACC__
 #include <cuda_fp16.h>
@@ -384,10 +386,12 @@ struct alignas(2) Half {
 #endif
 };
 
+#ifndef C10_EMBEDDED
 C10_API inline std::ostream& operator<<(std::ostream& out, const Half& value) {
   out << (float)value;
   return out;
 }
+#endif // C10_EMBEDDED
 
 } // namespace c10