Adapting llama_multiprocess to use `rsmpi`

Has anyone considered adapting `llama_multiprocess` to run on multiple machines instead of multiple processes? I've started by using the `SystemCommunicator` from `rsmpi` library to replace `nccl::Comm`, but the debugging seems pretty obscure. I managed to get it running but the last step hung forever. I can share more details.

Start with the llama_multiprocess entry point and trace how nccl::Comm is used alongside rsmpi's SystemCommunicator. Reproduce the reported final-step hang and use the available issue details to determine what is needed for llama_multiprocess to complete across multiple machines.

Written by the indexing model from the issue text.