Fix OSU latency benchmarks to actually measure latency

[dssgabriel]

Problem

The benchmarks in perf_tests/mpi/test_osu_latency.cpp do not implement OSU-style latency measurements. Each timed iteration performs only a one-way transfer: one rank sends while the other receives. Essentially, we are measuring the time spent by each

The OSU latency benchmark uses a ping-pong exchange and reports one-way latency as half the measured round-trip time (RTT). The current results, therefore, do not represent latency and cannot be meaningfully labeled as such.

Required Changes

• Implement a complete ping-pong in each timed iteration:
  • Rank 0 sends, then receives the reply.
  • Rank 1 receives, then sends the reply.
• Report one-way latency as half the round-trip time.
• Apply the same communication pattern to:
  • Kokkos Comm core API (MpiSpace)
  • S&R KokkosComm::mpi API
  • Raw blocking MPI baseline
  • Kokkos Comm core API (NcclSpace)
  • S&R KokkosComm::Experimental::nccl API
  • Raw NCCL baseline.
• Remove the nonblocking raw MPI baseline (to match what OSU benches actually implement).
• Fix the benchmark counters to reflect the communication performed.