fix: lumina 2 timesteps handling

[urlesistiana]

fix #2201

mainly replaced get_noisy_model_input_and_timesteps with the one from flux_train_utils.py.

This also added flux_shift timestep sampling method and --ip_noise_gamma support.

[duongve13112002]

@kohya-ss, Would you mind reviewing this and merging it into sd3? This issue could cause LoRA training or full fine-tuning to not work correctly.

[kohya-ss]

@duongve13112002 Sorry for the late confirmation. Is it correct to understand that this replaces #2215?

[duongve13112002]

@kohya-ss This implementation by @urlesistiana is more detailed and complete than the previous one I fixed. I’ve tested it and it works. I’m still fixing issues related to the full fine-tune of Lumina, and I’ll open a pull request once it’s done.