Running with multiple workers can produce `Sys_error("Invalid argument")` during flushes

[shonfeder]

I can reproduce this on my Windows VM by running the the parallel tests. What seems to happen is that some call to flush output raises a Sys_error("Invalid argument"). If I am not mistaken, this can happen either in the worker OR in the scheduler. When it happens in the worker it produces a deadlock, because the worker never returns from the select, and if in the scheduler the test will just fail.

Perhaps this is related to ocaml/ocaml#13586?

The current workaround is to force testo to run without workers by passing -j 0 in test invocations.

[mjambon]

Please provide a minimal example if you get the chance.

Meanwhile, we probably should harden the Testo code that handles channel redirections such that errors are reported precisely when a test function does something wrong such as closing stdout. I created a separate issue for this: #125

[mjambon]

First, we should set up CI checks for Windows: #120