This approach should make it easier to add new checks without having to
write lots of YAML and doing things against Ansible (e.g.
ignore_errors).

A single action plugin determines what checks to run per each host,
including arguments to the check. A check is implemented as a class with
a run method, with the same signature as an action plugin and module,
and is normally backed by a regular Ansible module.

Each check is implemented as a separate Python file. This allows whoever
adds a new check to focus solely in a single Python module, and
potentially an Ansible module within library/ too.

All checks are automatically loaded, and only active checks that are
requested by the playbook get executed.

Trying to improve the name, `init` needs to be loaded before calling other
subroles.

We don't make `init` a dependency of `common`, `masters` and `nodes` to
avoid running the relatively slow `openshift_facts` multiple times.

Allow reuse via role dependency.

We can use openshift_facts's service_type to determine the package
names.

Note: on a simple example run of ansible-playbook against a single
docker-based host, I saw the execution time jump from 7s to 17s. That's
unfortunate, but it is probably better to reuse openshift_facts, than to
come up with new variables.