Closes CUNY-CL#60.
Closes CUNY-CL#218.

LightningCLI removes our need to create separate training and prediction
CLI programs, moving nearly all of that logic into the base model.

This commit in particular sets the stage:

* Updates dependencies.
* Increments minor version number.
* Creates an empty `cli.py` where the CLI-speific logic will live.

This adds an extra flag to every command and for what? If you want to
keep model directories separate for different experiments in a way that
goes beyond already done with the automated versioning, you can just
append a subdirectory name to `--model_dir`, e.g., using `--model_dir
models/foo` instead of `--model_dir models --experiment foo`.

Borrowing a design element I used in UDTube, I decompose the dataset
object into two pieces:

* a `Mapper` interface which knows how to map between lists of strings
  and tensors (to decode and encode)
* `DataSet`, as before

There was no particular reason for the mapper functions to live inside
the dataset, and this commit simply makes this separation.

A subsequent commit will use this mapper object during prediction.