Performance optimization for dynamic setitem: Call op set_value to speed up because the original call to TensorToPyArray will introduce unnecessary data copy. (#30817)