Multiple new optimizations for SadTalker:

• A new option of size can be used to choose the model, (256 or 512) currently. eg, python inference.py --size 512
• A new 512x512 face render model is released, which may show better results in the month open and higher resolution.
• set new preprocess: extcrop and extfull, where the face crop is larger. eg, python inference.py --preprocess extcrop
• We package nearly all the models using safetensor and remove some necessary weights: checkpoint packages from 7GB -> 1GB!
• Reduce the requirement of installation, now, dlib or face-alignment is not a necessary package for a generation!
• Since we have not trained the new mappingnet, we still use two different mappingnet as the previous release. It will be merged into a unified safetensors in the future.
• add a script of webui.sh for Linux.
• Multiple webui functions are added: pose style, face model resolution, batch size, etc.
• Merge PR of #303 to boost the speed of face enhancers. Thanks for this contribution! very cool!