[CUDA][HIP] incorrect linkage of shadow variable for inline device variable

[yxsamliu]

C++17 inline variables have weak linkage in ELF files, which means multiple copies in different translation units are merged into one copy by the linker.

nvcc emits it as weak in both device and host:

https://godbolt.org/z/n8EjvWKxe

currently clang only emits it as weak in device but not host. This causes multiple shadow variables registered for the same device variable.

https://godbolt.org/z/Maz7aMhcP

This caused issue in HIP:

ROCm/hip#3463

[yxsamliu]

sorry I missed -fgpu-rdc when compiling using clang. when -fgpu-rdc is used, the linkage is weak for inline variables:

https://godbolt.org/z/K7vd8EGqa

So clang does not have the issue.