bugfix: improve CXI support for ALCF Aurora configuration

[ericjbohm]

Add support for 8 NICs as found on Aurora.

[lvkale]

This looks like a straightforward change, but since I have not looked at this code before:
does it handle only those 2 cases (1, 4 and 8)? I see the comment in code line 842 (near the assert) address this, but I am curious. I guess we will find out when the assert fails. What about cloud environments (or similar features coming on supercomputers) where a single node is allocated to multiple jobs, which each job presumably getting a subset of NICs.

[ericjbohm]

This looks like a straightforward change, but since I have not looked at this code before: does it handle only those 2 cases (1, 4 and 8)? I see the comment in code line 842 (near the assert) address this, but I am curious. I guess we will find out when the assert fails. What about cloud environments (or similar features coming on supercomputers) where a single node is allocated to multiple jobs, which each job presumably getting a subset of NICs.

The four NIC case is special because the optimal ordering on machines like Frontier is linked to both the GPU and NIC, and a simple in order mapping [0,1,2,3] will be suboptimal. So, really the 8 cpu case is one that could probably be handled by a general approach with a carve out for 4. But there is only one such known configuration.