[llvm] [offload] Split olMemAlloc to separate functions (PR #209196)

[Łukasz Plewa via llvm-commits]

lplewa wrote:

> Does the `olMemFree` still operate on these? 
Yes, it does. 
>Also, for AMDGPU at least, I believe host / GTT allocations still need a device to own the buffer, even if the memory is shared between many devices and resident on the host.
I'm not sure - I check a code and based what seen i do not see that device is required.


https://github.com/llvm/llvm-project/pull/209196