Add OpFunctor and replace cast, scale,clip, bce_loss and abs_grad with elementwise_no_broadcast

[AnnaTrainingG]

PR types

Others

PR changes

OPs

Describe

Add OpFunctor and replace cast, scale, full, clip, bce_loss and abs_grad with elementwise_no_broadcast
cast 当前在pten中修改不触发benchmark，现补充性能测试：

cast	case	type	old / us	new / us	speed up
case0	[16, 1785]	bool	1.328	1.413	0.94
case1	[16, 1]	int32->int64	1.27	1.239	1.03
case2	[16, 1, 513, 513]	int32->float32	43.09	42.229	1.02
case3	[30522, 1024]	float16->float32	246.57	236.28	1.04
case4	[1]	in64->float32	1.276	1.29	0.99
case5	[16, 16, 1]	float32->float16	1.269	1.312	0.97
case6	[16, 16, 1024]	float16->float32	2.025	1.842	1.10

case0 出现性能下降，主要原因是case规模比较小，机器波动影响占比较大。

[paddle-bot-old]

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

[MingMingShangTian]

LGTM for cast kernel

[MingMingShangTian]

LGTM for cast cuda kernel.

[JamesLim-sy]

I agree with this pr, but I think it better to code like this :

ScaleFunctor(InT scale_data, InT bias_data, bool is_bias_after_sacle) :
          : bias(bias_data), scale(scale_data), bias_after_scale(is_bias_after_sacle) {}