Move schema mutation orchestration into the mutation engine

The emitter used to walk the schema namespace itself, bucket types by kind,
dedup scalars, and run a second classification pass to split models into
input/output variants. That loop also leaked implementation details (an
`originalToMutated` back-map and a `TypeUsageResolver` that outlived
mutation) into the emitter's public shape.

Introduce `GraphQLMutationEngine.mutateSchema(schema, typeUsage)` that
returns a fully-classified `MutatedSchema`:

- Pre-bucketed models: `interfaces`, `outputModels`, `inputModels`
- Pre-classified operations: `queries`, `mutations`, `subscriptions`
- Derived metadata: `wrapperModels`, `scalarVariants`, `scalarSpecifications`

Classification happens inline during the walk. `typeUsage` is consumed
entirely inside the engine; the emitter hands it over and never touches
it again. `originalToMutated` no longer exists outside the engine.

Implementation lives in `src/mutation-engine/schema-mutator.ts` as a free
function called from the `GraphQLMutationEngine` class to avoid a circular
import between engine.ts and schema-mutator.ts.

The emitter's `emitSchema` reduces to: resolve type usage, call
`engine.mutateSchema`, report void-return diagnostics, check for empty
schema, build variant lookups, hand off to the (stubbed) renderer.

Author: FionaBronwen