This originally was a prerequisite for introducing some asymmetry (for finer-grained zombies) in the maps, which bimap made impossible because it guarantees 1:1 (even when using insert instead of insert_no_overwrite, what it does is it removes any entries related to either side first, then adds the new pair of entries).

But it turned into a refactor of its own, once I realized I can get rid of per-entry allocations, using the power of tcx.arena.dropless. _{(an untyped arena that allows allocating individual values, or whole slices, as long as they're !Drop - this works great because there is no "drop to remember", which reduces overhead and soundness complexity - and by using the slice feature, at the very least Vecs can be replaced, and even entire data-recursive trees could be arena-allocated, even if it sadly doesn't generalize to e.g. HashMap/BTreeMap)}

I haven't profiled this but it might be a bit faster? We were being pretty wasteful before, esp. lookup_type would .clone() variants of SpirvType that had Vecs in them, whereas now it's merely copying &[T]s.

Probably best to review each commit separately, matching the refactor stages.