[clang] [HIP] Use linker wrapper for device-only code object links (PR #211675)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

Ping. This came from an offload PGO Jira request, but it also affects device-side code coverage work in TheRock and device-side PGO through Comgr, so I think there is some urgency to unblock it.

The complexity mostly comes from an existing gap in the new offload driver: device-only compilation was not unified into the clang-offload-packager / clang-linker-wrapper pipeline. This patch does that, so structurally I think it is cleaner than adding more special handling in the AMDGPU device-only path.

The extra wrapper changes are needed because existing tools and apps use both bundled and unbundled device-only outputs. To preserve those workflows, the wrapper needs to support both output forms.

https://github.com/llvm/llvm-project/pull/211675