adding scripts and tests for naturalness metric

[sharan21]

PR for feature addition related to issue 129.

Motivation:
In Summarization, Generation, and Style Transfer Tasks, it is also useful to check the Fluency/ Coherence of the generated outputs (references in this case). The traditional method of using perplexity depends highly on the architecture of the pre-trained model chosen and does not correlate with human correlations as well as other newly introduced metrics. The "Naturalness" metric, for example, uses a Neural Classifier whose task is to determine whether a given sentence is a real English sentence or a fake. It is shown more to correlate with human judgments in determining the coherence of the generated outputs. This feature might be of use for people will similar needs.

Please refer to the full publication for more information: https://aclanthology.org/N19-1049/

Feature Addition:
The "Naturalness" metric has been added to scores variable in both functions calls and the NLGEval API. Unit tests to validate this metric has been added in test_nlgevals.py. All dependencies for this metric have been added to the nlgeval/naturalness directory, complying with the design architecture of the other metrics.

Behaviour:
The metric calculates the mean Naturalness score of the references provided. In case a list of references is provided (such as in NLGEval.compute_metrics()), the mean naturalness score of the union of all lists is calculated.

Please let me know your thoughts on this feature addition or if anything changes are required. Thanks!

[msftgits]

All CLA requirements met.

[temporaer]

Apologies for the long delay in response, @sharan21, and thank you for the contribution! Does the code and data you added come with some sort of license? I'd appreciate it if you could add that before merging. Thank you!