Support multiple single gpu's training.

[LEON-gittech]

It's understandable the current unsloth only supports 1 gpu's training. However, the current implementation hinders performing multigpu with multiprocess, like device 0 for process 0 and device 1 for process 1. I think that could also be regarded as single gpu's training. I hope your team will support this feature soon.