Unify Def and Type

[Hugobros3]

This (work-in-progress) PR makes Type a subtype of Def. This simplifies a fair bit and gets rid of some duplicated facilities in World and TypeTable. Additionally, the following adjustments were made:

• All Defs received a world argument and currently store it as a field. This enables pervasive sanity checking during node construction, to ensure all the operands come from the same world.
• Worlds are no longer mutated in-place through swapping their insides, instead a new Thorin class was created to contain the current world, and old worlds are discarded after importing. This is necessary to make the previous point work out, and is far less confusing to work with.
• The Importer is now supplied with the destination world, instead of owning it
• Nominal defs are now rebuilt using a pair of stub() and rebuild_from() virtual functions, rather than being special-cased in the Importer
• Params are now structural nodes
• Types operands are now defs, but a special types() helper is provided to access the type operands of struct types and such.
• Because all defs have a type(), a new special Star (universe) node was added, to serve as the type of type-level defs and itself.

• The World is now responsible for constructing and holding types - TypeTable still exists but only caches some of the simple types for speed.

  TODO:

  • Fix LLVM codegen issues
  • Unify/simplify the rewriting mechanisms (Importer vs Rewriter)

[Hugobros3]

Rodent works now, including on the GPU (modulo the existing RDNA2 issues). I think this means this branch is pretty close to regression-free...

The changes to get around structural params, warrant rewriting the free_defs() stuff too, as the old comment suggests.