Fix various bugs with LoRA Dreambooth and Dreambooth script

[patrickvonplaten]

This PR fixes the the issues #3279, #3346, #3363 and #3296. Here we also make sure that LoRA checkpointing is in a nice format. After this PR the format of LoRA checkpointing for dreambooth has the format:

├── optimizer.bin
├── pytorch_lora_weights.bin
├── random_states_0.pkl
├── scaler.pt
└── scheduler.bin

which allows one to easily load lora weights from checkpoints.

Script is tested for resuming from checkpointing for both single and multi GPU setups.