Add tf.keras model implementation

[eladn]

Main changes:

• Move to TensorFlow 2.0.0-beta1.
• New Keras model, implemented in keras_model.py. It uses tf.keras module (and not the pure Keras package). In order to use the new added Keras model, one should provide the --framework keras argument to code2vec.py. The default framework is still the old TensorFlow model (this is the one chosen if no additional arguments are stated).
• The keras model allows TensorBoard output while training.
• Move the TensorFlow model to tensorflow_model.py.
• Implement model_base.py for common implementation-independent model methods, from which both tensorflow_model.py and keras_model.py inherit.
• Use tf.Dataset for input pipeline and re-implement the reader in path_context_reader.py.
• Adapt the TensorFlow model to work with the new reader.
• Minor refactor the TensorFlow model.
• Move configurations from common.py to config.py.
• Added parameters and properties to config.py.
• Refactor the printing format of the model evaluation results.
• Use python's logger as output method instead of print(.) s.
• Added option to separate between <OOV> and <PAD> special words. Use parameter SEPARATE_OOV_AND_PAD in config.py to set whether to apply this option. The default is False (as it was used to be).
• Add requirements.txt file.
• Add python type-annotations to all parts of the code.
• The README.md has been updated accordingly.

What left to do:

• Train the new Keras for 8 iterations, upload it to S3 and add the link to README.

Should not break the current behavior, just added new functionalities. If new params not stated explicitly the default behavior haven't changed.