Working textual inversion with 6GB VRAM

[narnianpony]

According to https://rentry.org/sdupdates#newsfeed there is a new way to make TI work with 6gb vram.
I think it could be a very good feature to add to webui. Thanks!

Training a TI on 6gb https://pastebin.com/iFwvy5Gy 
 Have xformers enabled.
 This diff does 2 things.
            enables cross attention optimizations during TI training. Voldy disabled the optimizations during training because he said it gave him bad results. However, if you use the InvokeAI optimization or xformers after the xformers fix it does not give you bad results anymore.
            This saves around 1.5GB vram with xformers
            unloads vae from VRAM during training. This is done in hypernetworks, and idk why it wasn't in the code for TI. It doesn't break anything and doesn't make anything worse.
            This saves around .2 GB VRAM

        After you apply this, turn on Move VAE and CLIP to RAM and Use cross attention optimizations while training

[ClashSAN]

It's added #4056