Created by: typhoonzero

For reference: https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html

• update conv cudnn bwd for fp16
• more accurate mode for fp16 softmax (need future log_softmax support)
• The optimizer change is for loss and gradient scaling in: https://github.com/PaddlePaddle/models/pull/1533

TODO: support log_softmax and nll_loss (need test full training)

Previous work: https://github.com/PaddlePaddle/Paddle/pull/14992