return tensor instead of list of rnn and add support for multiple network in DNN

[jungle-cat]

Three modifications:

1. Return tensor of rnn operations while the original implementation tflearn 
    return list of tensors.
2. Add support for multiple network input in model.DNN, which is useful in 
    multi-task learning. The original implementation of DNN expects input 
    network to be a tensor which is hard to feed in multi-task learning.
3. Return tensor (single task/network) or list of tensors (multi task/network) in Evaluator.predict.

[jungle-cat]

Anyone can review this pull request?