fix topk big tensor bug

[liuruyan]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

修复 top k big tensor 问题，由于测试样例过多，因此未对全部失败样例进行测试，对两类错误进行了采样测试：

1. 精度错误：
   paddle.topk(Tensor([1, 1, 2281701379],"float32"), 9, axis=-1, largest=False, ) pass
   paddle.topk(Tensor([1, 1, 2281701379],"float32"), 13, axis=-1, largest=True, ) pass

2. cuda700
   paddle.topk(x=Tensor([8, 268435457],"float64"), k=4, ) pass
   paddle.topk(Tensor([6, 7, 51130564],"float64"), k=2, ) pass

修复 median big tensor 问题：

1. 精度错误：
   paddle.median(Tensor([3, 4, 190141782],"int32"), 0, True, ) pass
2. cuda700
   paddle.median(Tensor([3, 380283564, 2],"int32"), 0, True, )
3. 额外报错，可能为用例问题，需分析
   [paddle error] paddle.median(Tensor([1073741824, 4],"float16"), -1, False, )
   unsupported operand type(s) for *: 'Tensor' and 'Tensor'

pcard-67164

[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.

[wanghuancoder]

LGTM 同理，PR可以合入，但仍担心有路径覆盖不到，建议全量复测