Tf-encrypted

**Added**

- `tfe.local_computation` decorator, which is now the preferred way for providing inputs/outputs to a secure computation in TFE. All of the examples have been updated with example usage.
- `tfe.keras.layers.DepthwiseConv2D` -- converter support for the layer will land in the next release.
- More ops supported by the converter, including `SplitV` and `tf.keras.layers.BatchNormalization`
- `set_weights` for tfe.keras models and layers now accepts private variables as well as numpy arrays.
- `secure_model` now supports batch predictions.

**Changed**

- Examples now use the tfe.keras API for building models instead of lower level tfe Ops.
- Documentation is now generated from Google-style docstrings. As a result, we are only building docs for the tfe.keras API. Docstrings for other modules will be converted to Google-style and published progressively.

**Fixed**
- `tfe.keras` layers will now check to see if defaults have been changed from their originals in tf.keras, and surface errors whenever modified kwargs aren't supported. Some layers were failing to instantiate because these checks were too specific.

**Added**

- More `tfe.keras` functionality to convert `tf.keras` model into `tfe.keras`:
- `set_weights`
- `from_config`
- `model_from_config`
- `clone_model`
- Example with conversion of `tf.keras` model into `tfe.keras`.
- Improved handling for cases where the secure random operation is not available
- Added methods to the converter to inspect TF and Keras graphs
- `tfe.convert` now supports more than 2 inputs to `tfe.concat`

**Fixed**

- A bug in `tfe.keras.layers.Batchnorm` where the `offset` and `scale` conditions where inverted
- A bug in `tfe.convert` where ops with multiple outputs where not handled properly
- A bug in `tfe.convert` where it couldnt't convert a model correctly when there was more than
one special ops in the graph

**Added**

- More `tfe.keras` functionality, including BatchNormalization.
- `channels_last` support for `tfe.layers.Batchnorm`
- `tfe.convert` now supports conversions for Ops with multiple output tensors.

**Changed**

- All test files have been moved into the tf_encrypted namespace next to their corresponding functionality in line with the TensorFlow style guide.

**Fixed**

- A bug in `tfe.convert` where tf.split was only returning the first element of the output tensor array.
- An ImportError for users of TF 1.14+.
- A bug where `__radd__` and `__rsub__` were actually computing `__add__` and `__sub__`.
- An overzealous AssertionError in the tfe.serving.QueueServer.

**Added**

- More `tfe.keras` functionality, including `Sequential` model

**Added**

- First steps for the new `tfe.keras` module closely matching that of TensorFlow.
- More examples around private prediction and private training.
- Various notebooks, from getting started to in-depth debugging.
- Decoupled triple generation for Pond and SecureNN, allowing triples to be generated up front or simultaneously with other computations.

**Changed**

- All code is now following the style guide of TensorFlow.

Migration to third party organization, including on [GitHub](https://github.com/tf-encrypted/tf-encrypted/) and [Docker Hub](https://hub.docker.com/r/tfencrypted/tf-encrypted).