[compiler-rt] [compiler-rt] Add COMPILER_RT_GPU_THIN_LTO to build with -flto=thin (PR #226916)

[Wenju He via llvm-commits]

wenju-he wrote:

> the compiler and linker have to support one form of “object” linking

I don't think so. ThinLTO core idea is summary based thin-link for each TU and then parallel codegen. It doesn't mandate where the linking of binary objects happens. We do the thin-link/import step in clang-sycl-linker and do the linking in the NEO runtime. This is a working ThinLTO model.

https://github.com/llvm/llvm-project/pull/226916