A lot of this complexity was a consequence of trying to avoid redundant
extraction of the namespace's/contract's name, so on the other hand
letting it be redundant made all of the complexity kind of evaporate.
Add to that that we're now building a little deep list and then
flattening it, and this logic was able to get really neat in a way that
I couldn't work out a year ago.
Now we convert the ACI into trees of opaque types, then flatten the tree
into a map and a list of function specs, and only then dereference the
types in the function specs down to our accelerated annotated types.
Variants were working by accident, since
{variant, [{"VariantName", [Element]}]} had a similar enough form to
the opaque types that would come from something like
`type1(type2(int))`, but records were not working, since they have a
different form. Now both are handled explicitly so that only the
intended forms of each are handled.
