OpenMP GPU offload fails with NEURON's datareturn_py_gpu test when cell-permute=2

[pramodk]

Describe the issue

This is based on internal slack discussion here. In summary, what we see is that the datareturn_py test fails with OpenMP GPU offload and cell permute = 2. Strangely, the issue doesn't appear if caliper profiling is enabled.

To Reproduce

Steps to reproduce the behavior:

• build neuron+coreneuron+nmodl+openmp offload~caliper and run above mentioned test
• See failing pipeline with above settings: https://bbpgitlab.epfl.ch/hpc/coreneuron/-/jobs/437053

Expected behavior
OpenMP offload with cell-permute=2 should also pass the above mentioned test.

Logs
See above mentioned log: https://bbpgitlab.epfl.ch/hpc/coreneuron/-/jobs/437053

System (please complete the following information)

• OS: BB5 / RHEL 7.9
• Compiler: NVHPC 22.3
• Backend: GPU

Additional context

See slack discussion: https://bluebrainproject.slack.com/archives/C023T9MRC3B/p1668615314386289?thread_ts=1668551167.002749&cid=C023T9MRC3B