Type-indexed random value generation.

[athas]

The purpose of this PR is to introduce a combinator-based module for generating random values of compound types.

Because of restrictions related to size types, when generating things that involve arrays, you need to split the generation into two parts: generating the sizes, and then generating the values of that size. This can be seen in how the arr generator accepts a specific size.

The actual random number generation piggybacks on the existing generators for primitives.

Here is an example of how to use it to define a generator for random matrices:

module RNG = xorshift128plus
module G = mk_gen RNG

def random_matrix (size: i64) : [][]f64 =
  let rng = RNG.rng_from_seed [i32.i64 size, 1, 2, 3]
  let size = G.fmap (\x -> 1 + (i64.abs x) % size) G.i64
  let (rng, (n, m)) = G.generate rng (G.pair size size)
  let (_, M) = G.generate rng (G.arr (n * m) G.f64)
  in unflatten M

This generator produces random matrices of size up to size. It's easy enough to make it generate random matrices of exactly size, of course.

[athas]

@melsman What do you think of this design?