Simplify operator development by making build system more automatic

[typhoonzero]

It's easier for developers to develop operators if we make our building system more "automatic".

As in #3801, we don't need to write lines in pybind/CMakeLists.txt and operators/CMakeLists.txt, further more, we can just let users to implemet xxx_op.h, xxx_op.cc/cu, test_xxx_op.py that's all!

[luotao1]

We should remove USE_OP(XXX) in paddle/pybind/pybind.cc. We can auto generate USE_OP(XXX) into pybind.h file when building paddle/operators/CMakeLists.txt.

Followings things should be done before:

• As USE_OP(add_two)，we should rename add_op.cc/cu/h to add_two_op.cc/cu/h.
• As USE_CPU_ONLY_OP(gather) and USE_CPU_ONLY_OP(scatter), we should remove gather_op.cu and scatter_op.cu，thus, paddle/operators/CMakeLists.txt can auto set USE_CPU_ONLY_OP for ops that only has .cc file.
• As USE_OP_ITSELF(identity), we can move identiy_op out from scale_op, and new identity_op.cc/cu/h. Thus, USE_OP(identity) is ok.