Move to https://github.com/stickeritis/sticker
CI: Fix for changed dylib extension on Tensorflow on darwin
Fix compatibility with older Rust versions

See:
https://github.com/alexcrichton/flate2-rs/issues/207
Update dependencies
Bump version to 0.6.0
Add fix suggested by clippy
Mark trait objects with dyn

Rust 1.37.0 gives warnings for bare trait objects.
shell.nix: add Tensorflow GPU improvements

- Override arguments of libtensorflow_1_14_0.
- Bump danieldk package set for Tensorflow with compute capabilities
  3.5, 5.2, and 7.5.
Nix: bump danieldk package set for fixed TF 1.14 on Darwin
Fix byte encoding

tf.strings.unicode_decode decodes to unicode codepoints. So, any non-latin-1
character falls outside the range. This change introduces actual byte
encodings.

In some experiments, I found that with real byte representations, it helps
to have two layers in the byte RNN (presumably to combine characters with
multi-byte encodings), so this also changes the default hyperparameters.
Add support for RNN-based subword representations

When subword representations are enabled, sticker builds a string
tensor containing each form in the input. The forms are converted to
(truncated) byte representations in the Tensorflow graph. Each byte is
then mapped to a byte embedding. Subword representations are then
formed by applying a bidirectional RNN to the sequence of byte
embeddings.

In order to support the configuration of subword embeddings, the
sticker configuration format has been adjusted. An `input` section has
been added, which holds the new `subwords` boolean option. The
`embeddings` section is now a subsection of `input`. The use of
subword representations is communicated to the graph writers through
the shapes file.
Add references to papers that inspired sticker
Bump version to 0.5.1
Rename tf-proto to sticker-tf-proto
Improve encoder error handling
Restructure encoder modules

- The top-level depparse module is renamed to encoder::deprel.
- The Layer{Encoder,Decoder} are moved to the encoder::layer module.
CI: update to Tensorflow 1.14
CI: copy LICENSE.md and doc/ on deploy
Bump version to 0.5.0
Store graph construction args to the graph

Also provide the sticker-graph-metadata utility to print the args.
Next