Rebased version of #2215 with the following additions:

• I reviewed the code
• Deletes local state fields from caml_thread_struct that became redundant
• Frees local arenas in caml_free_stack (previously leaked for every fiber). This also applies to domains and systhreads.
• All effects tests are enabled
• Deleted CR about comparing stack->id; there's nothing to compare it with
• Changed XXX about the initial fiber stack size to a CR

If you run out of memory by creating too many fibers, you can get Fatal error: allocation failure during minor GC when the minor GC tries to malloc a block. This seems unfortunate.