[Request] Fix multi-gpu training for train_db.py and fine_tune.py

[SavvaI]

Can you please fix the multi-gpu training for train_db.py and fine_tune.py in the same manner it was done with train_network.py in the recent commit? For now it is not working as intended. I try to launch the accelerate launch --num_cpu_threads_per_process 1 train_db.py on my two GPUs with the same accelerate env as I am doing it with train_network.py #247 . But it turns out that when launching the train_db.py (and fine_tune.py) with accelerate it consumes a few times more gpu memory (or fails with not enough memory CUDS) per one gpu, which makes its application inexpedient. I think this is the matter of the great importance, because the quality of the trained model greatly depends on the effective batch size. Thank you.

[kohya-ss]

This issue has been fixed with #448. Please reopen if there is any issue.