[amdgpu] Implement D2D memcpy as HSA call

[JonChesterfield]

No description provided.

[JonChesterfield]

This works on a dual gfx906 system (copying over pcie) as written. Developed using a fork of a benchmark that can be found at https://github.com/JonChesterfield/multigrid-poisson which is not immediately appropriate for adding to aomp's test suite.

Note that functional tests are somewhat tricky in that they pass with and without this patch, all that changes is performance.

It's not dramatically quicker on my local system but might be on better hardware.

[ronlieb]

no longer applicable