Do not build sequential when asked for parallel
Right now we build the sequential version of the code even when asking for parallel version.
This was justified to allow the use for utilities when on the login nodes of cluster that do not allow MPI run of one proc on the frontal.
But this has become a lot of trouble with the use of third party libraries and not that usefull with the rise of GPUs (that are not always available on login nodes either (at least, not the same GPU as the compute nodes).
It is more convenient to allow one kind of build per build directory and ask explicitly the kinds we need.