Vak

Added
- warning when user runs `vak prep` with config.toml file that has a `[PREDICT]`
section *and* a `labelset` option in the `[PREP]` section.
- better error handling when parsing a config.toml file fails
+ traceback now ends with clear message about error parsing .toml file, but still
includes information from `toml` exception

Fixed
- tiny capitalization bug that broke configuration parsing

Fixed
- add missing sections and options to .toml file that is used to validate
user config.toml files, so that those options don't cause
invalid section / option errors

Fixed
- `vak predict` command now works for command line

Added
- [PREDICT] section now has `annot_format` option -- user can specify
whatever format they want, doesn't have to be same as training data
- [PREDICT] section of config now has `to_format_kwargs` option,
that lets user specify keyword arguments to `crowsetta.Transcriber.to_format`
method for the annotation format of files made from predictions

Fixed
- path in `PACKAGE_DATA` that captures 'valid.toml'

Removed
- `Dataset` class and related classes that were in `vak.dataset` sub-package
+ see `dataframe` module added below that replaces this abstraction
- dependency on `Tensorflow`
+ switch to `torch` because of consistent API, need to work with "mid-level"
abstractions, and preference for Python-first framework
- `core` sub-package
+ the idea is that the `cli` package should just implement all the logic that lets
people who don't want to program use the main functionality
+ and if you do want to program, the rest of the library should facilitate that
*instead of* trying to do all the work for you
- e.g. give someone w/basic coding skills friendly Python classes to work with
when writing a torch-vernacular training script, instead of
giving them a giant `train` function with 3k arguments that no one will ever use
- `AbstractVakModel` class -- gets replaced with `vak.Model` in `engine` sub-package,
see below

Changed
- `dataset` sub-package becomes `io` sub-package ("input-output", like in `astropy`)
- use `torch` and `torchvision` in place of `tensorflow`
- use `crowsetta` version 2.0
- switch to `toml` format for config files
+ more flexible than `ini` files, less code to maintain for parsing things that
don't fit into the `ini` format very well / not at all
- clean up `vak` package structure wherever possible: move many modules into
`util` sub-package

Added
- `dataframe` module in `vak.io`
+ essentially, data path is audio --> spect --> dataframe --> .csv file that represents
a dataset
+ choose to use external libraries that are already well-maintained and established to
handle as much of the data processing as possible, i.e. `pandas` + `dask`, instead of
trying to maintain a `Dataset` class that does all this work and deals with its own
filetype
- `datasets` sub-package
+ uses `torch` and `torchvision` abstractions to represent datasets + dataloaders
- `transforms` sub-package
+ uses `torchvision` transform abstraction to deal with things like "normalizing"
spectrograms
- `engine` sub-package
+ with `Model` class that models should sub-class; helps encourage consistent API for models
- `metrics` sub-package
+ to compute things like accuracy
+ lays groundwork for an `ignite.metrics` / Keras-like functionality

Fixed
- add missing line break in `installation.rst`
+ needed to show the crucial line about how to install from `--channel nickledave`