* remove in_dtype, out_dtype in redcue grad

* set the dtype and layout in noneedbufferInputs func

* move reduce_mean_grad kernel into phi

* move reduce_max/min_grad into phi

* remove raw max/min grad kernel

* fix bug

* fix max/min grad error

* move all reduce_grad kernel into one file

* add prod grad kernel

* add infermeta for prod kernel