Improve fortran type analysis with seeding

Part of the issue in https://github.com/EnzymeAD/Enzyme/issues/2966#issuecomment-5058930593 is that we struggle to discover types, we could use `enzyme_type` to emit types from the Fortran semantics

(might have to be careful around `TRANSFER`, `EQUIVALENCE` or any other method to exploit a type-hole)

Unsure where this could live and if we would need it to be part of the MLIR plugin in #2968 & co or if we could use it just on the Flang plugin level.

Start with issue #2966 and its referenced comment to understand the type-discovery problem, then investigate whether `enzyme_type` belongs in the Flang plugin or the MLIR plugin discussed in #2968. Account for the type-holes noted around `TRANSFER` and `EQUIVALENCE`; done means Fortran semantic types are reliably emitted without breaking those cases.

Written by the indexing model from the issue text.