【报bug测试2】多维Tensor梯度属性无法传播

[Ligoml]

bug复现环境（bug reproduction environment）

2.2.0.rc0-gpu 版本
单机单卡

bug复现步骤及最小代码集（Bug reproduction steps and minimal code set）

以下方式赋值时，两个输入的梯度信息不能同步到输出上

tiou[i:] = inter/union
使用 list.append + concatenate 方法速度很慢

import paddle
import time

a = paddle.rand(shape=[1,4])
b = paddle.rand(shape=[1,4])
a.stop_gradient = False
b.stop_gradient = False

print('=====paddle=====')
d = paddle.zeros((4, 4))
print(d.stop_gradient)
c = a/b
d[0, :] = a/b

print(a.stop_gradient)
print(b.stop_gradient)
print(c.stop_gradient)
print('Is d requires grad: ', not d.stop_gradient)

#---------------------------------------------
import torch
a = torch.rand([1,4])
b = torch.rand([1,4])
a.requires_grad = True
b.requires_grad = True

print('====torch====')
d = torch.zeros((4, 4))
print(d.requires_grad)
c = a/b
d[0, :] = a/b

print(a.requires_grad)
print(b.requires_grad)
print(c.requires_grad)
print('Is d requires grad:' ,d.requires_grad)

期望结果（Desired result）

=====paddle=====
True
False
False
False
Is d requires grad:  True
====torch====
False
True
True
True
Is d requires grad: True

实际结果（actual result）

=====paddle=====
True
False
False
False
Is d requires grad:  False
====torch====
False
True
True
True
Is d requires grad: True

paddle与pytorch行为不一致，stop_gradient属性不会传递

其他补充

No response