[Prim][CINN] Fix group_norm_grad decomp sum dtype

[lshpku]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

修复group_norm_grad组合算子拆解中sum的dtype设置错误的问题，即半精度下应当基于float类型进行sum，解决test_group_norm_op_deprecated单测的精度问题（这个单测在CI里不跑，只能手动验证）

参考layer_norm_grad：

Paddle/paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h

Lines 896 to 901 in 4f9d17a

	auto scale_grad_tmp = (x_sub_mean_mul_sqrt_var_1 * out_grad_cast)
	.sum(un_normalized_axis, x_cast.dtype(), true);
	scale_grad_tmp = reshape<T>(scale_grad_tmp, {-1});
	scale_grad_tmp = ConvertToOrig<T>(scale_grad_tmp, scale_ptr->dtype());
	set_output<T>(scale_grad_tmp, scale_grad);

检查了下另外几个batch_norm/layer_norm/instance_norm的反向都是对的，为啥只有group_norm_grad是错的

Pcard-85711

[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