Refactored for Python 3, breaking backwards compatibility

A considerable amount of refactoring occurred for this build, chiefly to get
rid of Python 2.7 support. While the functionality did not change much for this
version, we have switched from a pkgutil-style pydrobert namespace to
PEP-420-style namespaces. As a result, this package is not
backwards-compatible with previous pydrobert packages! Make sure that if any
of the following are installed, they exceed the following version thresholds:

• pydrobert-param >0.2.0
• pydrobert-kaldi >0.5.3
• pydrobert-speech >0.1.0

Miscellaneous other stuff:

• Type hints everywhere
• Shifted python source to src/
• Black-formatted remaining source
• Removed future dependency
• Shifted most of the configuration to setup.cfg, leaving only a shell
  in setup.py to remain compatible with Conda builds
• Added pyproject.toml for PEP 517.
• tox.ini for TOX testing
• Switched to AppVeyor for CI
• Added changelog :D

Bug fix from v0.2.0

This release is essentially identical to v0.2.0, modulo a bug fix.

Too many bits-and-bobs to count. I should keep a change log. Main features seem to be

• Optuna hooks for data loaders (see pydrobert-param)
• N-gram language modelling on the GPU. Can read in ARPA-LM files. Good for shallow fusion AM+LM.
• SpecAugment layer for, well, SpecAugment
• A variety of fixes and new features to my data loaders.

This is the last release that will support Python 2.7.

SpecAugment, Optuna, N-gram language modelling, and bug fixes

EDIT: Overnight, pytorch released a version that caused an error. Do not use this version! I'll update and release v0.2.1.

Too many bits-and-bobs to count. I should keep a change log. Main features seem to be

• Optuna hooks for data loaders (see pydrobert-param)
• N-gram language modelling on the GPU. Can read in ARPA-LM files. Good for shallow fusion AM+LM.
• SpecAugment layer for, well, SpecAugment
• A variety of fixes and new features to my data loaders.

More features! Attention, edit-distance based objectives, and more

Welp, the previous version should've been 0.1.0, but now we're at 0.1.0.

Besides bug fixes, this version includes

• The layers submodule, which currently includes attention mechansims for seq2seq. This also includes transformer network attention
• Convert from torch data dirs to NIST TRN and CTM, and back again
• Edit-distance based losses and reward functions, including Optimal Character Distillation
• Cleaner documentation on website, with tutorials

Lots of new stuff

Welp, I guess I never published the v0.0.1 tag... I just released it. What's new? Probably

• pydrobert.torch.data has been cleaned up. There are lots of tools for transducing data. Plus there are examples.
• pydrobert.torch.estimators has been added. The functions therein can be used for discretely sampling data. More examples.
• miscellaneous bug fixes

Enjoy