[unbox] Unbox Variants with Scalar Sets

[btwj]

This PR reworks how variants are solved. Instead of the previous approach using unifyRepresentations, the target provides a mapping from Type to a Scalar.set.

• Normalization of records now uses the SSA Interpreter, so logic isn't duplicated.

TODOs/Issues

• [ ] TODO: Heuristics for picking the scalar (if sets are equal, try using that first).
• [x] TODO: Don't regenerate IR for ClassAlloc for every record, do it once and cache.
• Issue: there is a problem with the JVM backend when doing IntEq with a long and 0:int (generated by zeroConst). Had to create new boxed zeros and ones.

[btwj]

@titzer I'm currently having trouble with the JVM target when it generates the heap image wrt closures. This happens when we have code pointers from a closure stored in a reference slot (#ref), see ub_cell00.v3. I'm also not that sure what the JVMV3ClosureAdapterGen is doing. Could you give me a couple pointers?

Scratch that, I figured out how to fix it.