Improve dialogue modeling

[kbalog]

Some particularly problematic parts:

• [x] next_action in dialogue_policy.py
• [x] generate_dact in nlu.py
• [ ] update_state_user in dialogue_state_tracker.py

For more robust dialogue modeling, the following solution was proposed. It considers not only the most likely user intent, but a ranked list of predicted user intents.

user_intents = predict_user_intent(user_utterance, conv_history)

for user_intent in user_intents:
  Check if user_intent can be executed  
    Do something
    Break
  Else
    Sorry, I don’t understand

[IKostric]

Depends on #107 and #65.