[Reland] Use static_assert to detect get_type_index used in device code (pytorch#139966)

pytorch#139173 was reverted due to an internal build break of using get_type_index in device code. This PR is created for ease of importing into META to further investigation.
Pull Request resolved: pytorch#139966
Approved by: https://github.com/malfet, https://github.com/huydhn

Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>

Author: 2 people

c10/util/TypeIndex.h

@@ -93,17 +93,14 @@ inline constexpr uint64_t type_index_impl() {
 
 template <typename T>
 inline constexpr type_index get_type_index() {
-#if !defined(__CUDA_ARCH__)
+#if defined(__CUDA_ARCH__)
+  static_assert(false && sizeof(T), " Don't call me from device code");
+#endif
   // To enforce that this is really computed at compile time, we pass the
   // type index through std::integral_constant.
   return type_index{std::integral_constant<
       uint64_t,
       detail::type_index_impl<std::decay_t<T>>()>::value};
-#else
-  // There's nothing in theory preventing us from running this on device code
-  // except for nvcc throwing a compiler error if we enable it.
-  return (abort(), type_index(0));
-#endif
 }
 
 #if !defined(TORCH_PEDANTIC)