Merge maps in OpRegistry and simplify register macros

[JiayiFeng]

In this PR, we fix #3435, and changed definitions of register macros:

// register an operator as well as its gradient version
REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)

// register an operator without any gradient version
REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)

And also there are some small alterations in USE_XXX macros:

// use an operator and all its available kernels
USE_OP(op_type)

// use an operator and its CPU kernels
USE_CPU_ONLY_OP(op_type)

// only use operator itself, no searching or linking job of its kernels
// e.g. RNN op
USE_OP_ITSELF(op_type)

Most operators have corresponding gradient operators, but in our code, some of them have not been completed. To pass compiling, I have to register these forward-only operators by REGISTER_OP_WITHOUT_GRADIENT for the moment. PLEASE AMEND THEM TO REGISTER_OP AFTER FINISHING THEIR GRADIENTS.

[wangkuiyi]

It looks to me that we need only one of REGISTER_OP_WITHOUT_GRADIENT and REGISTER_GRADIENT_OP.

Say, if we decide to take REGISTER_OP_WITHOUT_GRADIENT, we can still have REGISTER_GRADIENT_OP, but let REGISTER_OP other than the final user to call it.

[dzhwinter]

when we change the registration style, we need to discuss a previous problem. The user's code needs to write USE_OP(OP), USE_OP(OP_GRAD) twice. It is annoying.

[reyoung]

Just USE_OP(OP) is OK. because we always register forward operator and backward operator in same .cc/cu file.

USE_OP(OP) will force C++ linker to include all symbols in that .cc/cu file including forwarding operator and backwarding operator.

[dzhwinter]

we already have a unified registrar REGISTER_OP_WITHOUT_GRADIENT. I think we does not need this one any more.

[JiayiFeng]

You mean macro NO_GRADIENT(op_type) ?

[dzhwinter]

yes. Is it redundant?

[JiayiFeng]

Removed