Adjust spawn_process() to avoid unnecessary overhead processes: we can

just exec instead of creating a subprocess.  This reduces process usage
from four processes per parallel test to two.  I have no idea whether
a comparable optimization is possible or useful in the Windows port.