[clang] [clang][Driver] Honor `--offload-jobs` in device-only compilation (PR #233184)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

Thanks, Joseph. The Jira ticket reports this in HIPRT and Blender Cycles: they compile device-only code for many GPU targets into one bundled output. The old `-parallel-jobs` option made that work parallel, but `--offload-jobs` currently leaves it serial. Running one compiler per target works, but each build script would then need to manage those jobs and bundle the results. This patch keeps the existing one-command workflow fast.

https://github.com/llvm/llvm-project/pull/233184