Zookeeper

:bug: Bugs
- Correctly pass `is_training` to preprocess functions (46)

👷 Internal Improvements
- Test against TensorFlow `2.0.0-rc0` (45)
- Remove the `plot()` cli command (44)

:bug: Bugs

- Pass `data_dir` to `tfds.builder` 42

🎉 Features

- Allow users to overwrite `Preprocessing.__call__` (41)

:warning: Breaking Changes :warning:

- Change preprocessing to support arbitrary tasks (40)

🎉 Features
- Support preprocess functions that expect bytes (39)

👷 Internal Improvements
- Fix code coverage path (38)

🎉 Features
- Upgrade to `tensorflow-datasets1.1.0` (35)
- Remove experimental options (they will be enabled by default if we use the new S3 datasets e.g. `imagenet2012:5.0.0`.
- Reduce final prefetch to 1 batch, which should be enough for the current use cases and drastically reduces memory consumption
- Shuffle and cache before image decoding. This should reduce memory consumption of the shuffle and reduce the disk size needed when caching datasets. See https://github.com/tensorflow/datasets/blob/master/docs/decode.md
- Add support for dataset version using a `--dataset-version` option. See https://github.com/tensorflow/datasets/blob/master/docs/datasets_versioning.md

👷 Internal Improvements

- Add missing Python version requirement (33)
- Simplify CI config using `pytest-azurepipelines` (36)
- Add `pyflakes` and `black` linter (37)