Ml2p

------------------

* Improve batch prediction support to allow models to separately implement batch
prediction (e.g. a model might want to implement batch prediction separately to
improve performance).
* Tweak training job version format to only include major and minor versions numbers.
Patch version numbers are now reserved for models and intended for use in the case
where the code used to make predictions changes but the underlying model is the same.
* Model creation now defaults to using the training job with the same version as the model
but with the patch number removed.
* Endpoint creation now defaults to using the model with the same version as the endpoint.
* When creating training jobs or models, specifying the model type is now required if
the ml2p configuration file contains more than one model. If there is exactly one model
type listed, that is the default. If there are no model types, the docker file
must specify the model on the command line.
* Metadata returned by predictions now includes the ML2P version number.
* Version bumped to 0.1.0 now that versioning support is complete(-ish).

------------------

* Add support for client and server error exception handling.
* Deprecate passing a channel name to dataset_folder and add a new data_channel_folder
method to allow data in other channels to be accessed.
* Add dataset create and list commands to ml2p CLI.
* Add --version to ml2p and ml2p-docker CLIs.
* Allow model and endpoint version numbers to be multiple digits.

------------------

* Added validation of naming convention

------------------

* Added Sphinx requirements to build file.

------------------

* Cleaned up support for passing ML2P environment data into training jobs and
model deployments. Environment settings such as the S3 URL and the project name
are now passed into training jobs via hyperparameters and into model deployments
via model environment variables.
* Added support for training and serving multiple models using the same docker
image by optionally passing the model to use into training jobs and endpoint
deployments.
* Added support for rich hyperparameters. This sidesteps SageMaker API's limited
hyperparameter support (it only supports string values) by encoding any
JSON-compatible Python dictionary to a flattened formed and then decoding
it when it is read by the training job.
* Added skeleton for Sphinx documentation.
* Removed old pre-0.0.1 example files.

------------------

* Disabled direct internet access from notebooks by default.
* Added tests for cli_utils.