tests/perf: add shmem_barrier_perf latency benchmark

[bcmIntc]

Summary

• Adds shmem_barrier_perf to test/performance/shmem_perf_suite/ — a barrier latency benchmark that gives more diagnostic signal than a simple PE 0 stopwatch.
• Each PE times its own wait inside shmem_barrier_all(). Per-iteration max and min reductions across all PEs report the true slowest and fastest PE latency for each
  call. A sum reduction gives mean PE wait. All three are summarized as min/avg/max over all iterations.
• After the timing loop, shmem_double_fcollect gathers each PE's average latency to PE 0, which scans to identify and report the slowest and fastest PE by rank — useful
  for diagnosing load imbalance or NIC locality issues vs. collective algorithm overhead.

Test plan

• Build with and without --enable-hierarchical-barrier to confirm it compiles in both configurations
• Run on 2 PEs (smoke test): verify output is well-formed and min <= avg <= max for all rows
• Run at target scale (PPN >= 64, multiple nodes): confirm slowest/fastest PE ranks are reported and the gap between slowest-PE and mean-PE-wait rows is sensible
• Compare slowest-PE avg latency before and after enabling the hierarchical barrier to validate the improvement