Replace lod with recursive_sequence_lengths in book examples

* initial commit

* modify rnn_encoder_docoder example

In this way we can build and test using WITH_FLUID_ONLY flag being set
to ON.

- move paddle.v2.dataset,reader to paddle.dataset,reader
- remove unused code (which depends on v2) in paddle.dataset,reader

* Generated functions can take `*args` as inputs.

Use a new scope for inference in python unittest to avoid changing the value of variables for training.