This allows to parser more complex expressions in a single
functor (like the polynomial functor), instead of only unary and
binary parsers.

The downside is that implementing a functor parser got more complex:
Each functor now has to call the `inner` parser by itself and also
handle the case where the inner parser succeeds and the outer parser
fails separately.

This design doesn't sound particularly elegant, but there is no
obvious other solution.

This gets hairy pretty fast... In contrast, keeping the inner functor
type universal and using existentials only when needed allows to drop
in different existential type in different places (e.g tests).