Scalability improvements for disabled services

Currently, every service has a play in site.yml that is executed, and
the role is skipped if the service is disabled. This can be slow,
particularly with many hosts, since each play takes time to setup, and
evaluate.

This change creates various Ansible groups for hosts with services
enabled at the beginning of the playbook. If a service is disabled, this
new group will have no hosts, and the play for that service will be a
noop.

I have tested this on a laptop using an inventory with 12 hosts (each
pointing to my laptop via SSH), and a config file that disables every
service. Time taken to run 'kolla-ansible deploy':

Before change: 2m30s
After change: 0m14s

During development I also tried an approach using an 'include_role' task
for each service. This was not as good, taking 1m00s.

The downsides to this patch are that there is a large number of tasks at
the beginning of the playbook to perform the grouping, and every play
for a disabled service now outputs this warning message:

[WARNING]: Could not match supplied host pattern, ignoring: enable_foo_True

This is because if the service is disabled, there are no hosts in the
group. This seems like a reasonable tradeoff.

Change-Id: Ie56c270b26926f1f53a9582d451f4bb2457fbb67