instance_norm_grad kernel. Fix bugs that the data type of input is different from output in reduce_sum kernel. test=kunlun

d_bias are nullptr. Modify the code style of full_kernel.cc. Add new data
type for concat, elementwise_add, gather, scale, scatter ops. test=kunlun

* xpu unittest grad compute supports more types, *test=kunlun

* add instance norm xpu, *test=kunlun