improve CI testing to better handle changes to roles

[miabbott]

Our current methodology for testing PRs is to run any tests that are changed and if a role is changed, just run the improved-sanity-test. It's better than nothing, but less than ideal.

I propose we add some logic to .test_director to determine which roles have changed and run them individually in a dummy playbook. This would allow us to find dependency issues (because we are sloppy with the variable namespace and assume A LOT about what is available), but would also generally improve the kind of testing we are doing on PRs.

This could also grow pretty easily to determining which tests are utilizing the roles and running those tests as well.