Add an external option to toggle offload_to_cpu

[scissorstail]

https://github.com/huggingface/accelerate/blob/5565412840d1f4c215fe089aee733337f3d57d56/src/accelerate/accelerator.py#L3782C58-L3782C77

I was trying to fix an issue where training with FSDP on 2 GPUs hangs when saving the model at the end. While digging through the code, I accidentally discovered that setting offload_to_cpu to False makes everything work perfectly. Instead of modifying the library code directly, it would be great to have an option to configure this externally.

By the way, one really strange thing is that the training works fine the very first time after rebooting the computer, but from the second run onward, it always hangs

[S1ro1]

I think it doesn't hang (I thought so too on FSDP2) rather is insanely slow, can't do much with it. We can probably setup some external flag though, you're right