[Bug fix] Fix kaiming initializer div zero

[wanghuancoder]

PR types

Bug fixes

PR changes

Others

Describe

KaimingUniform初始化器，在如下代码下会发生除0问题：
paddle.nn.initializer.KaimingUniform has divide by 0 problem in following case:

import paddle
import numpy as np

paddle.enable_static()
x = paddle.static.data(name='x', shape=[1, 0, 0], dtype='float32')

kaiming = paddle.nn.initializer.KaimingUniform(0)
param_attr = paddle.ParamAttr(initializer=kaiming)

paddle.static.nn.prelu(x, 'all', param_attr=param_attr)

• 原因：这个初始化器，默认是使用Tensor shape的一个维度作为fan_in的，也可以由用户指定fan_in的值。但均不能为0。
• 解决方案：对这两种情况添加检查、报错。同时为这种场景添加单测case。
• Reason: This initializer uses one dimension of Tensor shape as fan_in by default, besides, users can also specify the value of fan_in. But both of them can not be 0.
• Solution: Add parameter check, error reporting, and unit-test for these two scenarios.

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.