[llvm] [Offload] Add OFFLOAD_FORCE_SYNC_OPS force-synchronization escape hatch (PR #222635)

[Jan Patrick Lehr via llvm-commits]

jplehr wrote:

> Doesn't CUDA have some environment variable for like CUDA_BLOCKING or something? Is that similar to this?

I think you're right. The internet told me `CUDA_LAUNCH_BLOCKING` exists and I think the behavior is quite similar.

There is also a subtle difference between this one and our downstream version: The downstream version switches the data transfers unconditionally to the 1-step synchronous copy, where as this one will still do the memcpy-staged transfer, I think. We could consider implementing what we have downstream, though it would be a little more intrusive.

https://github.com/llvm/llvm-project/pull/222635