add interpolation op

[luotao1]

solve #4319

[luotao1]

The construction of interp_grad op is :

operator.cc:142] Op(plain_net), inputs:{}, outputs:{}.
operator.cc:142] Op(plain_net), inputs:{}, outputs:{}.
operator.cc:142] Op(elementwise_add_grad), inputs:{Out[Out], Out@GRAD[Out@GRAD], X[MulOut], Y[Y]}, outputs:{X@GRAD[MulOut@GRAD], Y@GRAD[Y@GRAD]}.
operator.cc:142] Op(plain_net), inputs:{}, outputs:{}.
operator.cc:142] Op(elementwise_mul_grad), inputs:{Out[MulOut], Out@GRAD[MulOut@GRAD], X[MinusOut], Y[W]}, outputs:{X@GRAD[MinusOut@GRAD], Y@GRAD[W@GRAD]}.
operator.cc:142] Op(plain_net), inputs:{}, outputs:{}.
operator.cc:142] Op(elementwise_sub_grad), inputs:{Out[MinusOut], Out@GRAD[MinusOut@GRAD], X[X], Y[Y]}, outputs:{X@GRAD[X@GRAD], Y@GRAD[Y@GRAD]}.

[Xreki]

Good job! It is almost ok with some naming and description issues for me. About the op's name, I am not sure whether the abbreviation InterpOp is ambiguous.

[Xreki]

How about rename MinusOut to SubOut, since minus op is not used here anymore?

[luotao1]

Done.

[Xreki]

Would there more detail comments about inputs and outputs? About the format, you can refer to https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md#opprotomaker-names

[luotao1]

Done.

[luotao1]

About the op's name, I am not sure whether the abbreviation InterpOp is ambiguous.

Discussed with @qingqing01 , as numpy uses numpy.interp, we can use abbreviation name as well.

[luotao1]

Done.

[luotao1]

Done.

[Xreki]

LGTM