Federated fine-tuning recipes

[krammnic]

SmolLM was interesting example of federated paradigm for LLM applications. There are several frameworks for this task like OpenFedLLM, but they have different design problems. With simplicity and structure of torchtune it might be really interesting to propose something like this. I assume that just 2 new recipes is enough at this point:

local.yaml - recipe for local machines.
global.yaml - recipe for main global machine.

For each local recipe common procedure is done(same procedure as for one machine) on selected data partition. Then weights are aggregated by common FedOPT or FedProx procedure. @joecummings You had some ideas on this point either. Could you share them please?