I've updated my PR.

* After a discussion with Alexander Monakov (a GCC developer), moved vfork() into a small function to isolate it from both subprocess_fork_exec() and child_exec(). This appears to be the best strategy to avoid -Wclobbered false positives as well as potential compiler bugs.

* Got rid of VFORK_USABLE checks in function parameter lists. Now `child_sigmask` parameter is always present, but is NULL if vfork() is not available or usable. The type is `void *` to avoid hard dependence on sigset_t, which other CPython source files appear to avoid.

* Disabled vfork() in case setuid()/setgid()/setgroups() needed.

* Added more comments explaining vfork()-related stuff.

* Added commit message and NEWS entry.

Potential improvements:

* To avoid repeating long parameter lists in several functions, we can move them to a struct. The downside is that we'd need to convert child_exec() to use that struct all over the place. I don't have a strong preference here.

* Are any documentation changes needed? We don't change any interfaces, so I'm not sure.

Potential future directions:

* If desired, a vfork() opt-out may be implemented. But it'd need to disable posix_spawn() on Linux as well, since it might be implemented via vfork(), and we have no good way to check that.