More consistent dependency detection

[wlandau]

The problem

Because of the way automatic dependency detection works at the moment, drake workflows tend to fall out of date rather easily. It can be hard to tell which symbols drake detects and if dependencies were added or removed since the last make(), so a project's shelf life is shorter than it should be.

Proposed solution

We should put clear and firm boundaries on the scope of dependency detection. In the dependency graph, only keep nodes that are (1) targets in the plan, (2) objects in the environment, or (3) namespaced objects (referenced with :: or :::).

Drawbacks

Most workflows built with a previous version of drake will fall out of date. This was going to happen anyway because of #509, so it is really not so bad in this particular case. But in general, we should try to minimize the number of releases that invalidate old projects. So I have decided to

1. Postpone the upcoming CRAN release until the above changes are implemented, and
2. Name the next version 6.0.0 instead of 5.5.0.

Unfortunately, (1) means that @mschubert's great recent work on clustermq will not be factored into production drake as soon as I originally promised. Sorry about that, everyone.

[wlandau]

Fixed via #522. Unfortunately we still need codetools, so there is a little net gain in overhead. Makes #41 even more important.