Revert back to support input-hidden weights sharing between bi-directional RNNs.

1. Add options to enable and disable RNN weights sharing.
2. Set rnn_layer_size to 2048 by default.
3. Revert back the striding steps of 1st conv layer from 2 to 3.
4. Revert back to BRelu.

Above follows DS2 papers.

Fix a bug in use_gru argument parsing.

Print log to pfs for cloud training and set use_gru to False by default.

Add GRU support for DS2 model.

Rename self.local_data to self._local_data in class DataGenerator.

Fixed a serious error of bidirectional simple rnn for DS2.

metric ndcg with power 2

Fix a typo caused exception for audio_featurizer.py.

Fix dssm infer error with gru

deep speech2 can directly use warpctc instead by export LD_LIBRARY_PATH

Separate data uploading from job submission for DS2 cloud training and add support for multiple shards uploading.

Seperate data uploading from job summission for DS2 cloud training and add support for multiple shards uploading.

Bug fix and refine cloud training for DS2

Summary:
1. Add missing is_local argument (when set False, use pserver).
2. Add exception thrown if cp failed.
3. Add cloud mkdir if not cloud path for uploading does not exist.
4. Fix a bug using common path ./local_manifest for all nodes. (convert to /local_manifest)
5. Refine coding style.

Fix  local data variable in DS2 data generator.