add stableavatar inference

functionality for lipsync

add more force for i2v

fix endless logging

add i2v setting as default, and apply to integrated inference as well

integrated inference

add inference for different settings

small utilities

add audio warmup

improved embedding generation

fix reference input and teacher x0 pred

fix fa3 and add training validation

fixed audio and lora init. now identity at init

changes for checking functionality and debugging

working version of training and inference