Add Lin Domain.DLS test

[jmid]

This PR adds a quick Lin stress test of Domain.DLS.

It uses only one DLS.key for a start - not impressive but it is a start.
We might revisit it based on multiple-ts or consider replacing it by an STM test later.

Without the optional ~split_from_parent the child domain does not inherit the key, which makes for a fast negative test.
I did notice that the test takes longer and longer as it progresses, which may be due to the inability to remove DSL.keys again. For this reason, it seems faster to run the negative test first...
We might consider (a) running the property/test in child domain or (b) splitting the two tests to run in separate processes.

[shym]

Nice little test!
As keys cannot be freed and as new_key is really only called from init, it might make sense to use a constant key? 🤔 Or allocate a list of keys initially and pick one of them in init?

[jmid]

I grew increasingly unhappy with the short-comings the initial version, and therefore set out to write an STM test addressing some of them in ef4b42d.

The resulting STM test

• runs the two test properties in a child Domain (to avoid Domain.DLS polution across repetitions) and
• allocates an initial list of keys as suggested

I also adjusted the Util.repeat and retries parameters as a lot repetition didn't seem necessary to trigger it, and since I needed to write out the properties anyway to run the test properties in a child Domain.

This results in a negative test, that is quickly able to produce a counterexample where simultanious changes in two children are not seen in the sibling:

  Results incompatible with linearized model

                               |                           
                               |                           
            .------------------------------------.
            |                                    |                           
     Set (2, 1) : ()                      Set (0, 43) : ()                   
        Get 0 : 0                            Get 2 : 2            

[jmid]

A round of focused tests confirmed that the STM Domain.DLS test runs fast and works reasonable well.
I've therefore disabled the initial Lin test.

I expect this full CI run to fail its trunk workflows without #12545 merged into ocaml/ocaml.

[jmid]

CI summary

• all 17 trunk/5.2 workflows failed during opam install qcheck-core which has been reported and fixed upstream
• 2 MingW bytecode 5.1 timedout during threadomain [ocaml5-issue] Windows failures on threadomain #203

Out of 59 workflows and 19 failures, the 2 non-trunk failures were thus genuine.
I'll merge.