Support for Token Lengths Exceeding 75 Tokens in Text Encoder

[celll1]

Description:
This pull request introduces modifications to the Text Encoder to support input token lengths exceeding 75 tokens, specifically up to 3 chunks. The implementation has been confirmed to work with SDXL LoRA.

Key Changes:

• BOS/EOS tokens are now inserted every 75 tokens.
• An attention mask is applied to the intermediate special tokens.

Notes:

• Please note that the handling of BOS/EOS tokens differs from the current implementation in sd-scripts.
• I am uncertain whether applying the attention mask in this manner is the correct approach, and I welcome any feedback or suggestions for improvement.

Additional Private Modifications:
This PR also includes the following personal changes:

• Support for Adam-mini optimizer.
• Modifications to tqdm for use in Jupyter notebooks.
• Support for Hugging Face accelerator (distributed training has not yet been tested).
  Please note that due to these changes, the diffusers package needs to be updated to version 0.30.0 or later to enable resumption from a backup, as the model is sharded during backup.

I hope you find the necessary parts of this PR useful.