Add an output buffering for neuralmonkey-run

[varisd]

If I understand it correctly, when translating a dataset with neuralmonkey-run (using pretrained model), all the translations (or outputs) are held in the memory...
(this loop here: https://github.com/ufal/neuralmonkey/blob/master/neuralmonkey/learning_utils.py#L323)
... and only after producing all the outputs, they are being written into the output
(https://github.com/ufal/neuralmonkey/blob/master/neuralmonkey/learning_utils.py#L383)

When processing/translating large files, this might cause memory issues.

[jlibovicky]

Totally agree.

[jindrahelcl]

+1