Support HunyuanDiT finetuning all parameters

[Jarvis73]

An initial version of HunyuanDiT funetuning scripts. Only support save as ckpt.

The scripts support HunyuanDiT v1.1 and v1.2(default).

Differences between v1.1 and v1.2:

• V1.1 uses a noise schedule of scaled_linear(0.00085, 0.03), while v1.2 uses scaled_linear(0.00085, 0.018).
• V1.1 must enable --use_extra_cond for using image size condition and style condition, while v1.2 doesn't rely them any more (the PR's default behaviour).

Both versions must enable v_parameterization for training.

In a word, finetuning HunyuanDiT requires following arguments:

# Version 1.1
--use_extra_cond --beta_end 0.03 --v_parameterization --save_model_as ckpt

# Version 1.2
--v_parameterization --save_model_as ckpt

Finetuning CLIP and mT5 text encoder is not yet supported.