[inferno-core] Stop using Set for constraint accumulation, refactor infer a bit

[ngua]

There's no reason to use a Set for accumulating constraints:

• A Constraint is an Either (InfernoType, InfernoType, [TypeError SourcePos]) (Location SourcePos, TypeClass). Every Set operation had to compare deeply nested structures (walking entire type trees and error lists with source positions) to maintain ordering
• The constraint solver (processes constraints as a list via partitionEithers, and doesn't depend on uniqueness or ordering
• Constraints aren't duplicate in practice as they carry a [TypeError SourcePos] with unique source locations. Even if somehow there was a duplicate type constraint, the solver can handle it: after the first unification binds the type variable, the duplicate becomes e.g. Int ~ Int and succeeds (empty substitution)
• This is a basic accumulate-then-consume pattern; the asymptotics of a Seq are much better here than a Set

Some other changes:

• Because I touched many of the same locations, I decided to refactor the main inference function a bit. Instead of anonymous tuple projections, I created small helper records to improve readability
• These helpers will also be used when we do the rewrite from naive substitution-map approach to union-find for constraint solving

NOTE: Inferno tests pass themselves but need to test this elsewhere too