Another systhread rabbit-hole: caml_c_thread_(un)register. In order to make it better understandable, I refactored it in order to share code with thread_new, thread_start and thread_stop. Differences between thread_new+thread_start and c_thread_register are subtle and come down to when is the GC available to allocate the thread descriptor. We could share even more code if we change a few details, I can share some idea if someone has the time to implement it.

I think several bugs are fixed by this refactor:

• Fix TSan-reported race in tests/parallel/test_c_thread_register.ml #12773 : Missing restore (via @Johan511 & Fixes save_runtime_state logic in caml_thread_yield #12830)
• Fix calling c_thread_register without having initialized systhreads
• Fix invalid raising of exception and leak if tick thread is not created
• Fix missing memprof_delete_threads
• Fix situation where the main thread tries to stop before the registered C thread (?)
• Fix missing signal stack