Update example command lines in the module docstring of train_sft.py to remove weight_dtype=bfloat16 argument; this shouldn't be the default weight_dtype for SFT because of potential loss of gradient updates w/master weights in bfloat16. - #4710