[AST] NFC: Improve non-generic nominal type memory efficiency
The non-generic nominal type nodes do not actually need to use LLVM's
FoldingSetNode, and on my workstation the release build of the standard
library completes about 1/3 of a second faster after switching to LLVM
DenseMap. This is perhaps not surprising, because Decl to Type mappings
are only needed during early compiler stages, but the intrusive
FoldingSetNode data decreases CPU cache efficiency during all compiler
stages. As a bonus, the resulting code is simpler.
