feature request: dynamic memory allocation inside a gpu kernel

Creating a new issue on behalf of @Alkeryn...

https://github.com/intel/llvm/issues/879#issuecomment-560950093
> A thing cuda/hip has i wish opencl and sycl had tho is dynamic memory allocation inside a gpu kernel.

https://github.com/intel/llvm/issues/879#issuecomment-562076622
> Also i know you can exit an opencl kernel, have the host allocate more mem and enter again, but that's also kind of an hack, idk enough about gpu architecture to know if it would be possible but that would be neat to be able to allocate more memory without ever exiting kernel code / make a request to the host.