And leave only the type stuff without it.
This is because most expression stores don't have anything but types (e.g. generics, fields, signatures) so this saves a lot of memory.
This saves 58mb on `analysis-stats .`.
And make more queries non-interned.
Also flip the default for queries, now the default is to not intern and to intern a query you need to say `invoke_interned`.
This refactors how we deal with items in hir-def lowering.
- It now lowers all of them through an "ExpressionStore" (kind of a misnomer as this point) as their so called *Signatures.
- We now uniformly lower type AST into TypeRefs before type inference.
- Likewise, this moves macro expansion out of type inference, resulting in a single place where we do non-defmap macro expansion.
- Finally, this PR removes a lot of information from ItemTree, making the DefMap a lot less likely to be recomputed and have it only depend on actual early name resolution related information (not 100% true, we still have ADT fields in there but thats a follow up removal).
