Experiment: Generating OCaml instead of SML

[b-studios]

Just a draft for now.

While it is somewhat feature complete wrt to the old MLTon backend, there is still a lot to do:

• identify recursive groups and introduce let rec ... and ... binding groups for those (and those only).
• model capabilities / objects using records since this allows rank-2 types (https://soap.coffee/~lthms/posts/RankNTypesInOCaml.html)
• Implement heterogenous regions (since in OCaml we have Obj.magic)

[b-studios]

This was easier than expected.

@phischu if you are interested, you could benchmark this to compare it to our MLTon implementation used at https://github.com/effect-handlers/effect-handlers-bench

This way we could estimate how well the (global) optimizations performed by MLton perform compared to OCaml.

To get a first idea, here are the results of running examples/benchmark/nqueens on both:

hyperfine "./out/ocaml-main 12" "./out/mlton-main 12"
Benchmark 1: ./out/ocaml-main 12
  Time (mean ± σ):      4.070 s ±  0.002 s    [User: 4.060 s, System: 0.008 s]
  Range (min … max):    4.067 s …  4.075 s    10 runs

Benchmark 2: ./out/mlton-main 12
  Time (mean ± σ):     101.5 ms ±   0.5 ms    [User: 100.1 ms, System: 0.7 ms]
  Range (min … max):   100.7 ms … 103.0 ms    28 runs

Summary
  './out/mlton-main 12' ran
   40.10 ± 0.21 times faster than './out/ocaml-main 12'

Mlton is roughly 40x faster.

Using ocamlopt instead of ocamlc gives only a slowdown of 6x over mlton:

hyperfine "./out/ocaml-main 12" "./out/mlton-main 12"
Benchmark 1: ./out/ocaml-main 12
  Time (mean ± σ):     683.0 ms ±   1.9 ms    [User: 679.9 ms, System: 2.3 ms]
  Range (min … max):   681.7 ms … 687.5 ms    10 runs

Benchmark 2: ./out/mlton-main 12
  Time (mean ± σ):     101.2 ms ±   0.4 ms    [User: 100.0 ms, System: 0.7 ms]
  Range (min … max):   100.7 ms … 102.5 ms    28 runs

Summary
  './out/mlton-main 12' ran
    6.75 ± 0.03 times faster than './out/ocaml-main 12'