add leaky_relu forward and backward in activation_op.cu

[AnnaTrainingG]

PR types

Performance optimization

PR changes

OPs

Describe

PR说明：本次添加leaky_relu正反向向量化实现，并新增了REGISTER_ACTIVATION_GPU_KERNEL宏用于注册向量化OP，相比于原始REGISTER_ACTIVATION_CUDA_KERNEL 区别在于：ActivationGPUKernel（向量化实现）的调用；

leaky_relu新增向量化支持；
性能测试结果如下：

leaky backward	单位us
数据类型	pytorch	paddle old	paddle 1.8	paddle new	(new-pytorch)/pytorch	(new-old)/old	(new-1.8/1.8
float32	2.918	2.581	2.385	2.244	-23.10%	-13.06%	-5.91%
float32	113.29	117.43	116.37	114.33	0.92%	-2.64%	-1.75%
float32	2844.5	3142.6	3157.2	2864.7	0.71%	-8.84%	-9.26%
float32	287.74	298.92	296.15	289.8	0.72%	-3.05%	-2.14%
float16	2.236	6.173	2.193	1.823	-18.47%	-70.47%	-16.87%
float16	58.494	67.084	65.767	58.669	0.30%	-12.54%	-10.79%
float16	1438.8	4696.2	2158.5	1463.7	1.73%	-68.83%	-32.19%
float16	146.34	165.74	164.52	148.73	1.63%	-10.26%	-9.60%
leaky forward	单位us
数据类型	pytorch	paddle old	paddle 1.8	paddle new	(new-pytorch)/pytorch	(new-old)/old	(new-1.8/1.8
float32	2.846	2.49	1.824	1.834	-35.56%	-26.35%	0.55%
float32	77.749	78.571	81.747	77.263	-0.63%	-1.66%	-5.49%
float32	1956.2	2103.8	2338.1	1944.9	-0.58%	-7.55%	-16.82%
float32	197.6	201.29	210.73	196.41	-0.60%	-2.42%	-6.80%
float16	2.186	5.964	1.814	1.691	-22.64%	-71.65%	-6.78%
float16	47.231	48.692	47.756	41.399	-12.35%	-14.98%	-13.31%
float16	1115.4	1352.2	1509.1	1024.7	-8.13%	-24.22%	-32.10%
float16	116.08	120.26	120.52	104.09	-10.33%	-13.45%	-13.63%

在进行float16计算的时候，使用half22float2转换成float计算，是为了精度考虑，此外使用float2计算并不会导致性能下降，性能测试结果如下：

forward	half2	float2
[160800, 1, 7, 9]	75.977us	75.646us
[3, 513, 31, 31]	12.975us	12.469us
[8, 512, 157, 157]	734.99us	735.69us
[4, 128, 319, 31]	38.928us	39.265us
backward
[160800, 1, 7, 9]	53.284us	52.831us
[3, 513, 31, 31]	8.9540us	8.6850us
[8, 512, 157, 157]	512.87us	512.91us
[4, 128, 319, 31]	26.944us	27.039us

[paddle-bot-old]

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

[Xreki]

LGTM

[zhangting2020]

LGTM