Expand on stats to guide improvements

[jmid]

We have an initial stat setup in src/statistics which has already been handy to get initial Thread modes running.

Since we have issues with Lin and STM Thread-mode predictability and STM_domain.agree_prop_par_asym's usage of Semaphore.Binary not always triggering bugs, it would be welcome to expand on that stats setup to guide us towards stats-backed improvements.

Right now, one has to rewrite code to setup a stats experiment every time.
Ideally, we would have stats tests in place ready to run. Running them continuously on the CI would probably be overkill and a waste of CPU cycles - it would nevertheless be useful to be able to do so, to help make sure improvements are indeed improvements across across the various platforms and architectures.

It could furthermore back changes to the Lin/STM core such as #324 by being able to say, e.g., "with 95% confidence this change does not affect the bug finding rate".

An ideal setup would measure

• the fraction of failed properties (no shrinking required) of some large number of iterations, e.g., 37/20000
• across the usual ref and CList tests + a few more representable examples (Hashtbl?, ...) that also allocate
• test both Lin and STM properties in Domain and Thread modes

We should be able to do so, without too much new code or duplication by factoring the 'test specifications' into a reusable module akin to src/neg_tests/stm_tests_spec_ref.ml that can then be used from both the "regular test drivers" and from a "stats driver".

The src/neg_tests/stats.ml file of #99 contains steps in this direction.
This is a prerequisite for making progress on #338 (and #47) AFAICS.