Fix a race condition in SwappedCtx parallel exec

Fix a race condition in SwappedCtx parallel exec

The protection in set_maestro was failing to detect between the real
maestro and the working threads of the parmal. This is now fixed
according to helgrind.

Also, refactor things a bit by moving this set_maestro call directly
into the SwappedCtx constructor. Much less of a shotgun surgery :)