Secure multitenancy in engine->agent communication

Agent configuration is injected by the engine at deploy time. At the moment a shared username, password and vhost are used for all environments and instances.

Input queue names to resources are determined as a function of their object and environment id, so the chances of accidental collision is low.
Output queue names belong to an environment and currently are `-execution-results-<environment name>` (e.g. `-execution-results-quick-env1`). This makes collisions possible between environments owned by different users (or even the same user in some circumstances). (Filed bug https://bugs.launchpad.net/murano/+bug/1321476)

In addition to the chance of accidental collision, there's an opportunity for intentional mischief since nothing stops an agent on one host sending messages to another, presuming it is able to determine the id (as a proof of concept, an agent template could be constructed that queried rabbitmq for queues using the python API, for example). Queues need to be safe at least within tenants.

The output queue name problem can be made less problematic by using the environment id in the queue name (https://bugs.launchpad.net/murano/+bug/1321476) but we need a solution for authenticating agent queues.

Some suggestions that came up before:
* distinct username/password (per tenant)
* auth keys (per tenant)
* separate vhosts (per tenant)

The issue with all of these is that it means the engine needs a lot of access to rabbitmq. Marconi potentially would solve all these problems, but likely will not be guaranteed available in the juno timeframe.