Integrate new OptimizerInBackward in rest of torchtune codebase

[joecummings]

Context

We landed support for a better, cleaner optimizer in backward implementation in #2719! This was implemented only for full_finetune_single_device.py.

To-dos

Feel free to grab one or more of these, just comment on the Issue if you'd like to work on it, and thank you so much for helping contribute to torchtune :)

• Integrate OptimizerInBackward for PPO and QAT single device recipes.
• Integrate OptimizerInBackward into full_finetune_distributed.py (This should be taken on by someone familiar with the library, not a beginner)
• Integrate OptimizerInBackward into all distributed recipes that currently use it - needs the above to land
• Deprecate all uses of the hooks used previously to set up an optimizer in backwards (e.g. create_optim_in_bwd_wrapper

[omkar-334]

i'd like to work on this @joecummings

[krammnic]

@omkar-334 It might be tricky at some points (that's why no orange label). But in case if you are sure that it will not cause any struggles I would love to review your PR.

[krammnic]

Will take this, but ETA is ~1 week