Update to tensorflow v2 (use tf.keras)

[brollb]

This will require substantial updates to the parsing library and we will need to check if this will be a breaking change.

[umesh-timalsina]

Few notes on parsing to tensorflow.keras sources (v2.2.0):

1. keras.initializers for v2.0 is replicated from init_ops_v2.py
2. Preprocessing Layers added.
3. Non uniform import of super classes and inheritance: somtimes Layer is directly imported while at other times base_layer.Layer is used to create a Layer subclass.

[brollb]

Here are my thoughts:

• It would be good to follow imports from the initializers.py file to init_ops_v2 and init_ops. It will probably be a little annoying but will ensure that we actually parse the definitions for these initializers (which used to always be in the initializers.py file).
• It would be good to add them here too. It isn't necessary to add them in this PR though.
• It's not clear where this will become an issue. What is your concern here?

[umesh-timalsina]

1. Initializers can be parsed from init_ops_v2 since we are planning to migrate to tf > 2.0.
2. These layers are exported as experimental, but yeah, I agree, this should be added here.
3. Skulpt returns a different bases Array on one vs other. So, a lot of functions would either have to be refactored/ take into account the classes with seemingly same Super class, will yield different kinds of base Array.

Here is an example:
https://github.com/tensorflow/tensorflow/blob/2b96f3662bd776e277f86997659e61046b56c315/tensorflow/python/keras/layers/recurrent_v2.py#L68

If you replace this line with just GRUCell and change the import statement below before feeding it to skulpt, it will be uniform across the board.
https://github.com/tensorflow/tensorflow/blob/v2.2.0/tensorflow/python/keras/layers/recurrent_v2.py#L32

Using the file's content as is: The bases array for this file's AST for the GRUCell will be different not have an id.v attribute used heavily to determine base classes in our layer-parser.

[brollb]

Serialization should be easier with tf v2 :) https://keras.io/api/models/model_saving_apis/

[umesh-timalsina]

Tensorflow v2.3.0 was released yesterday. Keras section of the repo has changed.

[umesh-timalsina]

As discussed we are switching to python from skulpt for parsing. See #180.