AMD Suggestion, TunableOP tuning script

[Charmandrigo]

As suggest, this is to accelerate trainings on AMD cards that support TunableOP (such as 7900XTX)
https://pytorch.org/docs/stable/cuda.tunable.html

Accelerate could come with a prepared script to tune itself the first time it's configured, right now I had to manually run a tuning training, adding multiple tunableop env variables and using a small dataset, it doubled down my training speeds (for reference, cut down a 1024,1024 batch 8 sdxl lora training from 7.5s/it to 3.2s/it all thanks to the tunableOP tuning).
Could be extremely useful for RDNA3 owners!