[PyTorch] s/__attribute__((__noinline__))/__attribute__((noinline))/ (#52362)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52362

AFAICT, it is documented to be the latter and not the former.
GCC: https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#Common-Function-Attributes
Clang: https://clang.llvm.org/docs/AttributeReference.html

Both versions work in the oldest and newest GCC & Clang versions on Godbolt: https://godbolt.org/z/s6f4PW

So why change?
1) lack of underscores matches the documentation
2) AMD HIP defines `__noinline__` as a macro, which doesn't play well with the underscore version.
See 2080cc113a/include/hip/hcc_detail/host_defines.h (L54)
ghstack-source-id: 121875424

Test Plan: Rely on existing CI

Reviewed By: bhosmer

Differential Revision: D26488991

fbshipit-source-id: 6cfcdfd41c58170659e263cd519ac5359ffd5d46

c10/macros/Macros.h

@@ -179,7 +179,7 @@ namespace at { namespace cuda { using namespace c10::hip; }}
 /// C10_NOINLINE - Functions whose declaration is annotated with this will not
 /// be inlined.
 #ifdef __GNUC__
-#define C10_NOINLINE __attribute__((__noinline__))
+#define C10_NOINLINE __attribute__((noinline))
 #elif _MSC_VER
 #define C10_NOINLINE __declspec(noinline)
 #else