More helpful metadata

[space-nuko]

Some more metadata that could be useful to have later:

• Dataset/reg image dirs - the number of images in each directory, the name of each directory and the repeat count
• random session ID - uniquely identifies a set of models as belonging to the same training session
• keep tokens - another command line argument
• training date - date that the set this model came from was trained on, useful for historical purposes
• output name - the output name passed to the script

Let me know if there's more that could be useful

[kohya-ss]

Thank you for PR! It looks good!

I would like to add a feature to LoRA module to extend the support to Conv2d with 3x3 filter. It is possible to know if Conv2d is supported by the model, by the weights in safetensors/ckpt file, but I would like to have the version in the metadata as well.
(like "1.0.0" or "2.0.0" as semantic versioning, or simply "linear" and "conv2d".)

Therefore, I will merge this as soon as I finish adding that feature. If it takes too long, I'll merge this first.