This was a bug in the `delmin` helper that is used to delete the smallest
element of the right tree. It deleted too much.

This caused erroneous calls to rotateDoubleRight instead of rotateSingleRight,
thus disbalancing the tree and possibly failing to pattern match.

This is used in the Eq instance and also makes debugging and testing easier.

This is not strictly required by copar but allows for easier testing.

This returns the size of the tree, not the actual number of elements (there can
be more elements per node) and thus should not be in the public API.

This is a requirement for its use in copar and also allows for testing, since
`shouldBe` from hspec requires those two instances.

Due to the way the functor expression is parsed, this led to two nested
polynomial functors when functor rewriting is disabled. Without the parens we
get just one polynomial functor.

Alright, compromise. Word16 might have been a little too small

This should require 4 times less RAM and allow us to generate larger DFAs.

This allows to disable the 1-elem-block optimization to be able to compare the
running times with and with out it.

Queue.empty was being called with negative sizes

The same principle as when splitting blocks applies: The largest doesn't
have to be in the queue, because we already refine with respect to it
when refining the second-to-last block.

Checks if all weights are between 0 and 1.

Switches from GADT syntax to a normal ADT. This is more easily presentable in
writing as it doesn't require additional explanation and is shorter.

Although less explicit about the kind of `f`, it is more clear and easier to
replicate in publications.

This fixes up the haskell examples, so that tangling them to a file results in a
useable powerset implementation.