Add dropout module

[tatHi]

For the training of NN, it is known that dropout improves the performance.
I added the dropout module to the embedding, the LSTM output, and LSTM layers.
Also, I revised argparse to take the dropout rate, so now you can specify the dropout rate as follows:

$ python -m bi_lstm_crf corpus_dir --model_dir model --dropout 0.3

All dropout modules share the same dropout rate in this commit.
On the local NER dataset, this dropout contributes to +5% F1 improvement :)