Further investigate all-to-all performance on NVL72

[pentschev]

The performance we observe on NVL72 is still suboptimal, with NCCL being roughly 2x faster. We are currently at ~20GiB/s, where NCCL achieves ~40GiB/s on 64 GPUs.

We need to further investigate where we are losing. Most obvious probable candidates are:

1. RapidsMPF bench_comm
2. UCXX lock contention
3. UCX internals (could be a variety of different reasons)

Possible ways to explore this further:

1. All-to-all test with OSU+UCX
2. Properly configure to use MPI communicator with UCX backend on NVL72 (currently hitting issues with MLX5, probably due to minor misconfigurations that simply need to be corrected)