Fix cuda kernel launch of grid sampler

[wanghaoshuang]

PR types

Bug fixes

PR changes

OPs

Describe

Fix cuda kernel launch of grid sampler

Fix #29066

[paddle-bot-old]

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

[jerrywgz]

建议增加一个输入较大的单测case

[wanghaoshuang]

@Avin0323

关于benchmark CI failing的说明

对于inputs:

grid (Variable) - dtype: float32, shape: [4, 256, 246, 2]
x (Variable) - dtype: float32, shape: [4, 1, 128, 128]

在该PR之前， block_size = 4256246 / 512, grid_size = 512, block_size超过了max_thread_num, 所以计算结果实际是错误的。
在该PR修改之后， block_size = 512, grid_size=4256246 / 512, 计算结果正确，benchmark耗时增加。

benchmark部分超时log如下：

2021-05-26 13:02:59 [check_op_benchmark_result.py:80] [INFO] ------ OP: grid_sample_3 (forward) ------
2021-05-26 13:02:59 [check_op_benchmark_result.py:82] [INFO] GPU time change: 7.08968% (develop: 0.0168862 -> PR: 0.0180833)
2021-05-26 13:02:59 [check_op_benchmark_result.py:84] [INFO] Total time change: -2.24423% (develop: 0.0382667 -> PR: 0.0374079)
2021-05-26 13:02:59 [check_op_benchmark_result.py:85] [INFO] backward: False
2021-05-26 13:02:59 [check_op_benchmark_result.py:86] [INFO] parameters:
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	grid (Variable) - dtype: float32, shape: [4, 256, 246, 2]
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	x (Variable) - dtype: float32, shape: [4, 1, 128, 128]
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	align_corners (bool): False
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	mode (string): bilinear
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	out_shape (list): [4, 1, 256, 256]
2021-05-26 13:02:59 [check_op_benchmark_result.py:88] [INFO] 	padding_mode (string): reflection

[zhangting2020]

1. large input导致慢的原因是单测框架的期望梯度算的较慢吧？有测过大概需要多久吗
2. 如果已经用了skip_check_grad_ci，下面259～260就不需要写了

[wanghaoshuang]

1. 是的，单测框架的期望梯度算的较慢。单个case跑了20min，还没有完成。
2. 已删除259~260

[zhangting2020]

LGTM for skip_check_grad_ci

[jerrywgz]

LGTM