Data.Semigroup is in Prelude as of GHC 8.4. Since we require this
compiler version anyway, we might as well remove the Semigroup
imports.

This ensures hspec-discover is installed for test compilation. This is
unnecessary if the hspec-discover executable is already installed by
the build tool (stack), but that's not always the case.

Instead of just in the root directory

Those were broken by the recent data structure optimization.

Instead of an unlabeled tuple

This speeds up the polynomial functor by a factor of 4-5 for some of
my tests and also reduces memory usage considerably by using much more
efficient data structures for `H1 Polynomial`, `Weight Polynomial` and
`H3 Polynomial` (e.g bit-arrays instead of a recursive data structure
filled with `Bool`s).

This is most noticeable for DFAs which are essentially `X^A`, where A
is the input alphabet and can be large.

This is still work in progress. At the moment, only graphviz is
available as output format and the command line interface isn't final
yet.

We previously did not distinguish actions properly by using the
powerset functor for action transitions. By using P(AxX) where A is
the set of possible actions, we correctly distinguish different
action states that behave the same.

The main reason being that we need a newer criterion version to pull
in fixes for bugs that affected our benchmarks and the easiest way to
get that was to bump the stack resolver which incidentally also
upgrades the compiler.

The performance implications of this change are:

  - Microbenchmarks stay more or less the same, parsing benchmarks
    seem to get a little bit slower.
  - Real-World benchmarks on PRISM examples seem to get a bit faster
    overall.

This makes the interface easier to comprehend and write tests for.
There is now an extra function `addBlocksToQueue` that handles the
queue stuff.

These are both useful for testing.