Improved representation for Set and Map

[meooow25]

Consider changing the representation to avoid Tips.

Suggested by wrengr in #1069 (comment)

---

I'm thinking

data Set a
  = Bin !Int !a !(Set a) !(Set a)
  | Two !a !a
  | One !a
  | Nil -- Invariant: Never a child of Bin

This hopefully improves performance but certainly improves memory usage. Compared to today, for n elements we can avoid storing ~n Tips and ~n/2-2n/3 Ints.

Alternately, as midway between current and the above,

data Set a
  = Bin !Int !a !(Set a) !(Set a)
  | One !a
  | Tip
-- Invariant: Bin _ x Tip Tip is always replaced with One x

This puts the match-on-singleton idea of #1069 into the type. Avoids storing ~2n/3-n Tips and ~n/3-n/2 Ints.

[meooow25]

My suggestion to evaluate these:

• Adopt the most important functions (say fromList, insert, delete, member/lookup, union, intersection) to work on these alternatives in a separate repo. This is to avoid having to change everything if you try to edit Data.Set.Internal or Data.Map.Internal instead.
• Adopt property tests, to be safe. No need to write tests separately for each structure, CPP can toggle the import.
• Adopt benchmarks for the implemented functions.
• Run the benchmarks and compare the results to see which representation is better.

I can test this, but it is a decent amount of work and it will be a few weeks before I can start.

@HuwCampbell might you be interested in trying this out, as an extension of #1069?