This allows files to exists in the examples directory that don't have
a ".out" file. These are now ignored in the test suite. Such files
could be documentation or ephemeral tests.

... just to have uniform nomenclature

- Add API docs
- Make data fields strict and (if possible) unpacked
- Rename a few things for more consistency

Allow to use the unicode codepoint "LETTER CAPITAL P WITH HOOK" as
powerset functor symbol.

All files in examples/ now have a corresponding .out-file that is
checked against the algorithm's output on that example as part of the
test-suite.

'#' starts a line comment and skips the rest of the line.

We need this to print the final result with the names defined in the
input file instead of the internal node representations (ints).

Use parseErrorPretty instead of show

This was just there to ease the transition into the new functors
framework. Now that's done, it can be removed.

Since we now have all blocks in the same partition, we can finally fix
the type of split to be more intuitively related to the printed
version.

Instead of one big Algorithm.hs, there are now separate modules for
`initialize`, `split` and the queue handling stuff. This hopefully
makes the code structure a little more clear and also allows us to
export utility functions (like `initialize` and `split`) from
submodules for testing, while still keeping a tidy export list in the
toplevel module.

If needed, Show should be required at a different place.
RefinementInterface should now represent exactly what's described in
the paper.

Still not used by the algorithm, but at a step in the right
direction...

This hooks the new machinery into the old algorithm, but it's still
not finished.

- Sorts are currenly silently ignored when constructing the initial
  partitions (which should lead to wrong results).
- We do a little dance to feed the algorithm that expects a Vector of
  encodings a single encoding for a desorted functor expression.
- Examples in the source should be converted to the new format.

This changes the transformer stack from `ParserT State` to `StateT
Parser` for these reasons:

1. This allows the state to backtrack with the parser. While this
   shouldn't be necessary for the morphism parser, it's definitely
   less surprising behavior.
2. We need to compose the functor expression and morphism parsers.

Instead of expecting `ParserT` to be at the top of the monad stack,
they now only assume that the monad satisfies the `MonadParser`
constraint, allowing to use them unlifted in `StateT Parser`