RFC: major rewrite for world age and possibly inference?

[timholy]

Recent and pending PRs fix nearly all the test failures in JuliaInterpreter; the few that are left seem a consequence of the binding partitioning/world age on globals. I also took a peek at LoweredCodeUtils, the next stop on the march towards Revise, and from what I can tell, the bleeding-edge JuliaInterpreter fixes everything except about 7 test failures, all of which appear to be about world age. Thus I think this is the main issue left to address.

There seem to be two potential paths:

1. Extend support for world ages in the current framework. There is a start on that, but we'd have to ensure that all accesses to globals use invoke_in_world. We'd also likely want new options, e.g., @interpret world=:latest foo(args...). With this we cannot resolve globals during the "frame compilation" phase, we have to leave them as GlobalRefs.
2. Do 1, plus provide an invalidation framework. I suspect that the only possible approach here is to incorporate type-inferred code, and keep a reference to the corresponding CodeInstance; then, when you descend into a frame, you first check world ages to see whether it's been invalidated in the meantime. This would, I think, allow us to resolve globals during compilation, and obviously optimize our method dispatches as well.

In the short term, I strongly favor 1. 2 might be better done as part of a larger restructuring around, say, JuliaLowering. I'd be curious whether others have different thoughts about the right approach.