[openmp] Enable tests on new devicertl on amdgpu

[JonChesterfield]

@ronl we could land this without the fix to the compiler ABI. It'll break some set of tests on CI (which will probably be different to the ones it breaks for me locally), and we can then disable those tests, which hopefully leaves some still executing. That doesn't get us any closer to working but it does slightly stage the change over to the new runtime in terms of accumulated error data.

[JonChesterfield]

• rebase

[JonChesterfield]

I've checked in the three unsupported lines in tests so that enabling the new runtime by default (which would obsolete this patch) with a working compiler backend should pass CI.

[JonChesterfield]

• rebase

[JonChesterfield]

I'm told the compiler backend bug that blocked this is fixed, but clang has picked up some noise to force use of the old runtime which probably conflicts with landing this as-is. Going to abandon this patch instead of try to extend it into one that reverts workarounds in clang.

[JonChesterfield]

This is OK as originally written if D115150 lands first
edit: also fine without it, since the test setup of default + new is right for amdgpu at present

[pdhaliwal]

LGTM

openmp/libomptarget/plugins/amdgpu/CMakeLists.txt

 get_property(AMDGPU_ARCH_COMMAND TARGET amdgpu-arch PROPERTY LOCATION)
 
 execute_process(COMMAND ${AMDGPU_ARCH_COMMAND} RESULT_VARIABLE amdgpu_arch_result
                                                OUTPUT_VARIABLE amdgpu_arch_output)
 if (${amdgpu_arch_result})
   libomptarget_say("Not generating amdgcn test targets as amdgpu-arch exited with ${amdgpu_arch_result}")
 else()
   # Report to the parent scope that we are building a plugin for amdgpu
-  set(LIBOMPTARGET_SYSTEM_TARGETS "${LIBOMPTARGET_SYSTEM_TARGETS} amdgcn-amd-amdhsa" PARENT_SCOPE)
+  set(LIBOMPTARGET_SYSTEM_TARGETS "${LIBOMPTARGET_SYSTEM_TARGETS} amdgcn-amd-amdhsa amdgcn-amd-amdhsa-newRTL" PARENT_SCOPE)
 endif()