[clang] [OpenMP][Clang] Force use of `num_teams` and `thread_limit` for bare kernel (PR #68373)

[Johannes Doerfert via cfe-commits]

jdoerfert wrote:

I think the follow up, to force the user bound for bare kernels, make sense. I am not sold on this patch though. Why would we disallow users to do the same looping we do in the deviceRTL while hoping the offload runtime will pick a good grid size?

https://github.com/llvm/llvm-project/pull/68373