auto-round

fix critic bug for gradient_accumulate_steps!=1 and reduce cpu memory of lm-head tuning

#97

Merged