fine-tune update

[paul-lupu]

This PR is made to help others struggling with training on lower VRAM. I got this to work on 24gb on the RTX-4090 with the settings i provided in the readme.

Added a new export_pth() function that save the weights to an outputs/pth_export/<step_number>.pth file; this is to make it easier to test in actual comfyui workflows

Also I changed the save_checkpoint() function as follows:

• the next_save_iter step is now used to call the export_pth() function
• added a new parameter called --model_checkpoint_n_steps; every time the save_checkpoint function is called, it checks if the current step is past the value in this parameter and if so, also saves the full model