* add gelu composite rule

* use full replace fill_constant

* change the form of calculation

* remove float16 test for composite gelu

* reformate code

* remove float16 test case

* add forwad with prim and backward without prim test

* add float16 test for composite gelu and add high dims test

* add float16 test case and high dims test

* shield float16 and cpu test case

* increase train step to 10 in test cinn prim gelu

* replace pow to multiply