Created by: Jie-Fang

This feature implements mixed precision training on rcnn. It utilizes auto loss scaling, black white lists to make training stable. Testing this feature requires this pull request in Paddle and turns on --use_fp16=True and use_dynamic_loss_scaling=True. You can also set --loss_scaling, --incr_every_n_steps, --decr_every_n_nan_or_inf, --incr_ratio, --decr_ratio to control the detail.