Consolidated error handling when GPU only tests execute on CPU systems

[abouteiller]

Fixes issue #641

• [ ] capture ci tag and compile for Cuda or HIP based on this tag, not on availability of the software stack.
• [ ] decide if we keep the check in stage_main about counting GPUs. We now will detect (later) erroneous behavior and produce a parsec_fatal so maybe that is already clear enough?
• [x] cleanup error cases for when hook==NULL, cases have been split between when the dyld function was not populated (goto next_chore) and other cases (this is a bug, assert/fail).
• [ ] stress tester is not symmetrical https://github.com/ICLDisco/parsec/issues/641#issuecomment-2040819481