coverage metric: fraction of resolved global bindings

[StefanKarpinski]

See https://groups.google.com/forum/#!topic/julia-users/sk8Gxq7ws3w. In particular:

You can completely prevent any surprises if you have tests that exercise every global binding that occurs anywhere in your code. That's a fairly straightforward coverage metric and one that's not unreasonable to insist on keeping at 100% in production-quality code. It is much easier to satisfy than perfect test coverage: if you have many functions that use foo you only need one test that calls foo somehow to be sure that foo is an unambiguous reference.

Follow-up to #4345.

[StefanKarpinski]

Cool idea, unclear if it belongs in the language or an external package.

[vtjnash]

I think the expectation is it may require codegen support to mark each global as used when it's reached.

Though inference may be able to help give the proper counterfactual from an external package: mark as failing any statement that could be reached, but which appears it might throw an UndefVarError if it does.