Improve the compilation pipeline (introduce stages)

For now, we have somewhat messed up pipeline: the IR nodes do all by themselves. Type checking, member resolve, lowering, codegen: we have everything everywhere. Right now, only parsing is properly isolated from everything else.

We should introduce more formal stages of compilation and _maybe_ even create several different layers of IR? For example, we have certain constructs that have to be lowered (such as `+=`-styled operators). We may get rid of these constructs in the "lowered IR layer" and thus decrease the amount of type checks required in the codegen.

I don't know how should it be organized, though. Just start from separating the IR layers, and everything else will click and fit in place?

Thoughts?

cc @kant2002, @impworks

When implementing, look for number `201` in the source and try to eliminate every instance of that number.