Tweetynet

Changed
- translate config files in `./src/configs` to `vak 0.3.0` format
- specify minimum required version of `vak` as 0.3.0

Added
- add Levenshtein distance and segment error rate to metrics
[37](https://github.com/yardencsGitHub/tweetynet/pull/37)

Fixed
- change optimizer back to Adam, should not have been SGD
[36](https://github.com/yardencsGitHub/tweetynet/pull/36)

Added
- add `logger` parameter to `TweetyNetModel.from_config` class method
[34](https://github.com/yardencsGitHub/tweetynet/pull/34)
+ to take advantage of `logger` attribute being added to `vak.Model` -- i.e. be able to log training

Changed
- make `yarden2annot` return labels as strings [32](https://github.com/yardencsGitHub/tweetynet/pull/32)
+ to be consistent with what `vak` expects

Added
- `network.py` module (see "Changed")
- initial `tests/test_tweetynet.py` module, written for `pytest`

Changed
- convert to `PyTorch` model
+ because Tensorflow 1.0 is deprecated.
+ chose PyTorch over Tensorflow 2.0 for several reasons, see
[NickleDave/vak88](https://github.com/NickleDave/vak/pull/88)
+ network is now implemented as `TweetyNet` class in `network.py`,
as a subclass of `torch.nn.Module`
- make `vak` a dependency, since we use `vak.Model`
- rewrite `model.py` with `TweetyNetModel`, a `vak.Model` subclass
+ that specifies optimizer, loss, and metrics for `TweetyNetModel`
- `yarden2seq` re-written to work with `crowsetta 2.0`, renamed `yarden2annot`