expose max-autotune in configs for better perf

[felipemello1]

TLDR: when compiling, max-autotune will improve perf, but will take longer to compile. If the user has a long training job to do, it is definetely worth it.

TODO:

1. Define the flag in just one config
2. Find all places where we compile (prob: flex attention utils, compile model, compile loss)
3. Run a test without max-autotune and log with metric_logger=torchtune.training.metric_logging.WandBLogger
4. Run a test with max-autotune
5. Share results in a PR
6. If accepted, implement it for every config/recipe (ideally this should be implemented in the utility level, not recipe level)

tune run full_finetune_single_device --config llama3_2/1B_full_single_device dataset.packed=True tokenizer.max_seq_len=4096 dataset.split=train[:5%] metric_logger=torchtune.training.metric_logging.WandBLogger 

[krammnic]

@felipemello1 Let me clarify: we consider 1 model and 1 recipe, assuming that it is generalizing on all recipes and models, and then just use this mode across all compiles? And if it is accepted, expose and support it across all recipes. Sounds as an amazing task, will open PR.

[felipemello1]

@krammnic perfect. The idea is the do the least amount of work, and only after approved, make it work for everything. Thank you so much! looking forward to the PR