* Chage `IndicateDataType` to `GetKernelType`. Make it easier to
  understand.
* Change `OpKernelKey` to `OpKernelType`
* Make operator developers can customize which kernel the operator will
  use in runtime.

1. user can disable peephole connections.
2. not calculate some gradients.

Use fixed activation in the lstm kernel, since there is some bug in the activation function pointer. It will be fixed later.