Kur

- Lots more PyTorch bug fixes and support (including support for CTC loss).
- More custom Jinja2 filters/functions (e.g., 'load_json', 'gpu_count',
'ternary').
- Additional Kurfile expressions (e.g., 'when', reverse sort) and better name
inference for named layers.
- More layers (e.g., 'squeeze') and better layers (e.g., 'border' support for
pooling layers, highway dense, highway CNN).
- `speech_recognition` supplier supports multiple keys and does a better job
at guessing if bad data has been passed to it.
- Plugin support.
- Better GPU usage manipulation.
- Automatic model reuse to reduce memory footprint.
- Better shape-query utilities in the Model class.

- Custom Jinja2 filters and better scope parsing during Kurfile loading.
- Kurfile globbing.
- Time-based stopping criteria.
- Fixed serious batch normalization issue.
- Multi-validation/testing support

- Fixed a bug when processing meta-container inputs.
- Added globbing patterns for includes.

- Fixed a bug in the Keras parallelizing which crashed TensorFlow while waiting
for compilation to finish).
- Improved debug output for Keras 2.0.2.

- Very small update prior to the Deep Learning Hackathon
- Simplifies the requirements for subclassing Container
- Added some documentation about the text hook.
- Added an initial KurHub hook to support the hackathon.

- Improved GPU selection
- Documentation updates
- Better JSONL loading
- Couple minor bug fixes
- New layer: for_each
- Added templating and meta-containers