Saving results from each process and aggregating distributed output

[IlyasMoutawwakil]

The way results are handled with distributed inference/training (torchrun launcher) is by only saving them in rank 0, however this might be wrong, misleading or ambiguous, especially when processes are not necessarily synchronized (dp/tp inference) and where one process/device might be affected by comms. A better approach would be sending the results from each process and merging them appropriately by the launcher (sum throughputs, avg latencies ?). this will remove the world_size logic from benchmarks.

[IlyasMoutawwakil]

done in #122