Support zoneout on lstm cell state and add recurrent dropout

[azraelkuan]

hi，any plan about these two questions

1. lstm zoneout on cell state the same with hidden state
   

   haste/frameworks/pytorch/layer_norm_lstm.py

   Lines 62 to 66 in 9da2454

   	if zoneout_prob:
   	if training:
   	h[-1] = (h[-1] - h[-2]) * zoneout_mask[t] + h[-2]
   	else:
   	h[-1] = zoneout_prob * h[-2] + (1 - zoneout_prob) * h[-1]

2. add recurrent dropout the same with keras
   https://github.com/tensorflow/tensorflow/blob/fcc4b966f1265f466e82617020af93670141b009/tensorflow/python/keras/layers/recurrent.py#L2450-L2459

thanks!

[sharvil]

Do you have a specific use-case in mind for these two additions? I'm open to adding these features – it would be nice to have a sense of the scope and value of these changes.

[azraelkuan]

@sharvil oh，thanks to your reply.
for the first question:
in the original paper, it has two different zonout on lstm states.
for the second question:
the dropout will be benifit for very long sequence. like tts acoustic modeling