[Dy2Stat]Support Mixed Precision training in @to_static

[Aurelius84]

PR types

New features

PR changes

Others

Describe

[Dy2Stat]Support Mixed Precision training in @to_static.

Examle:

# suppose the forward has beed decorated by @to_static
model = MNIST()

adam = AdamOptimizer(
            learning_rate=0.001, parameter_list=mnist.parameters())

scaler = paddle.amp.GradScaler(init_loss_scaling=1024)

for batch in batch_iter_reader:
    with paddle.amp.auto_cast():
            prediction, acc, avg_loss = mnist(img, label=label)

    scaled = scaler.scale(avg_loss)
    scaled.backward()
    scaler.minimize(adam, scaled)
    mnist.clear_gradients()

1. 此 PR 为主体功能PR，后续会添加更多模型单测验证和完善@to_static 支持混合精度训练的功能

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhhsplendid]

LGTM

[zhangbo9674]

LGTM

[chenwhql]

LGTM for framework.py