Return non-zero if rosdep check cannot locate dependent

[SubaruArai]

This fixes #937.

As for the tests, I concluded it'll be quite difficult to implement with the current architecture.
The reason being the data in ./test/tree is expected to succeed, see here.

Please tell me if there's a good way to inject a non-functioning package, I'll add a test.

[cottsay]

This seems like the right call. An unresolveable rule is arguably a more critical error than an uninstalled package.

[tfoote]

Catching these errors seems like a good idea. We may want to add a command line flag to keep the old permissive behavior to allow someone who was relying on it to keep using it.

[SubaruArai]

@cottsay @tfoote I've changed the code, please review.
Somehow github insists there's still a change requested - please ignore it.

[tfoote]

That review status doesn't go away until. The review has been dismissed by a maintainer or changed by the reviewer.

This is very simple now and will avoid what used to be false negatives.

[cottsay]

We may want to add a command line flag to keep the old permissive behavior to allow someone who was relying on it to keep using it.

I think that providing a unique return code gives a developer the same ability to ignore the new behavior. Either way they'd need to change how they call the tool, but the unique return code is more widely usable than an explicit command line flag.