fix bn grad compute when x.stop_gradient=True

[ceci3]

PR types

Bug fixes

PR changes

OPs

Describe

fix bn grad compute when x.stop_gradient=True

import paddle
paddle.enable_static()

x = paddle.static.data(shape=[4,3, 10, 10], dtype='float32', name='x')
x.stop_gradient = True

bn = paddle.nn.BatchNorm2D(3)
y = bn(x)

avg_cost = paddle.mean(y)
sgd_optimizer = paddle.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
print(paddle.static.default_main_program())```

[paddle-bot-old]

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

[qingqing01]

Need to add UT.

[zhiqiu]

LGTM for ShareDataWith

[zhangting2020]

LGTM for skip_check_grad_ci