[llvm] [openmp] [Offload] Remove handling for device memory pool (PR #163629)

[Joseph Huber via llvm-commits]

jhuber6 wrote:

> Well, I'm not sure what our policy regarding using offload with libc. Is that required now? If not, it is okay for me to have our own implementation. Alternatively, even if we want to remove it, we want to document it somewhere that, if anyone wants to use device side dynamic allocation, they need to enable `libc` project; otherwise this will not work.
> 
> On the other hand, I think even our device runtime requires the support for dynamic allocation for the smart stack when static shared memory allocation runs out. This PR will make it completely not work.

AOMP shipped their own malloc, CUDA uses their own malloc, upstream libc has real malloc, this was never a functional implementation and just made a gigantic bump pointer to cover up the fact that we didn't have it. `libc` isn't "required", but if you have an application that needs more than a trivial bump pointer for malloc then you should use `libc`.

https://github.com/llvm/llvm-project/pull/163629