Instead of running valgrind on each job, half of which are in release
mode and don't have much usable information for valgrind, perform an
debug build as part of allowed_failures and run valgrind on that one,
which should speed up the feedback we get from the builds.