Enable batch norm and bidirectional LSTM

[ASvyatkovskiy]

This PR enables small modifications to the model builder:

1. Add Keras bidirectional wrapper to LSTM
2. Batch normalization between linear and non-linear layers. Namely, it would work like this:

   pre_rnn_1D = Convolution1D(num_conv_filters//div_fac,size_conv_filters,padding='valid') (pre_rnn_1D)
   if use_batch_norm:  pre_rnn_1D = BatchNormalization()(pre_rnn_1D)
   pre_rnn_1D = Activation('relu')(pre_rnn_1D)

   Both features are disabled by default and can be controlled in the conf.yaml

3. Add some comments in the code