Implementation of CUDA device id selection (--device-id 0/1/2)

generalized some functions and option for ignoring first layer

added support for extensions

Preprocess: fixed keep ratio and changed split behavior

Features to readme

add refresh button for aesthetic embeddings
add aesthetic params to images' infotext

turns out LayerNorm also has weight and bias and needs to be pre-multiplied and trained for hypernets

Training UI Changes