Add examples for scaling LLM inference on ALCF systems

[rickybalin]

Add scripts showing how to scale LLM inference on Aurora and Polaris via vLLM. Included are example with MPI, EnsembleLauncher and Dragon covering a few different types of workflows.

To Do:

• MPI on Aurora
• MPI on Polaris
• EL batched on Aurora
• EL batched on Polaris
• EL request-driven on Aurora
• EL request-driven on Polaris
• Dragon request-driven on Aurora
• Dragon request-driven on Polaris
• Performance plot comparing approaches on Aurora
• README with instructions for all approaches

Improvements:

• Do not do explicit batching of the prompts for the MPI case.
• In MPI script, don't scatter prompts from 0, can read them from local-rank and scatter within node group or directly read from all ranks since prompt file is on /tmp.
• In MPI script, likely don't need the HF token as argument, it's already set/checked in submit script.
• CPU bindings in EL batched case.
• vLLM initialization time does not scale well on Polaris for both MPI and EL_batched approaches.