Make Pruning TF 2.X Ready

[alanchiao]

Subsequently update https://www.tensorflow.org/model_optimization/guide/pruning to indicate compatibility.

Progress: submitted commits that are relevant:

• ff464f9
• f13568b
• ab96106
• dbae704
• 10fd9f8
• 105ec70
• 029972d

As a byproduct of this (by updating calls to TF to tf. instead of from ... import), we will also:

• ensure we don't mix 1.X and 2.X behavior (we currently do)
• prevent end user breakages when using the pip package
  • For instance, we used the tf.keras.experimental.load_saved_model and tf.keras.experimental.save_saved_model APIs, but the path switched from "from tensorflow.python.keras.saving import saved_model" to "from tensorflow.python.keras.saving import saved_model_experimental", making the code no longer compatible with older TF releases. By switching to just "tf.keras.experimental.load_saved_model", this would have been prevented. This can happen in general since TF doesn't guarantee the file paths don't change, only the public APIs won't change.
• improve maintainability

Remaining issues:

• Consider use case of loading 1.X model into 2.X code.
• Consider case of importing SavedModel. Our whitelist doesn't include the SavedModel versions of the layers. E.g. "<class 'tensorflow.python.keras.saving.saved_model.load.ZeroPadding2D'>" isn't the same as tf.keras.layers.ZeroPadding2D. Note that SavedModel no longer uses its versions of the layers, as of TF nightly due to tensorflow/tensorflow@568427a

[alanchiao]

This is considered complete now and will be released in 0.2.1. We are still doing some additional model training (e.g. mobilenet converges to right accuracy) to validate, but there is high confidence that it's ready for use.