* add pure fp16 training.

* update the usage of cast_parameters_to_fp16.

* use fp32 to initilize parameter values and then cast to fp16.

* enable multi_precision when training with fp16.

* add static loss scaling.

* add the rescale_grad function in the pure fp16 training.

* use cast to complete the pure fp16 training.

* use sum loss for fp16 trainging and mean loss for fp32 training.

* refine codes according reviewers' suggestions