* Restore --config. This will be useful when you have an init config
that you don't want overwritten.
* Cache the individual transformed images in TextualInversionDataset.
This gains speed by avoiding reading and reprocessing the image each
time it's used for training.
* Turn on no_grad for inference and clean up tensors during
checkpointing. This reduces memory usage slightly.
* Set the sample output size to 384x384. We just need them large enough
for manual evaluation, and this gains us a decent chunk of speed.
* (breaking change) Custom templates are now semicolon-delineated.
Additionally, custom templates are properly passed through to
TextualInversionDataset to generate input_ids for your images. Using
custom templates which accurately describe your input images seems to
improve training fidelity.
* Cache autoencoding of image pixel data. This substantially increases
the speed of training, upwards of 40% for me.
* Clean up a little bit of cruft.
