Introduce self-describing hooks

[ruudk]

@hook(input: ["creator.id"]) forced every call site to know and select the exact fields a hook reads internally. Updating a hook's needs meant touching every query that used it — the call site was coupled to the hook's implementation.

A hook now declares the data it needs as a named GraphQL fragment via #[Hook(name: ..., requires: ...)]. The call site shrinks to a bare @hook(name: ...). The generator builds a typed data class from the requires fragment (emitted into Generated/Hook/), validates that the @hook field sits on a type the fragment's on Type condition allows — an object type, or an interface, in which case any implementer — and injects the required selection into the operation sent to the server. The hook receives the typed object.

Fields a hook needs but the caller did not select are merged into the raw payload only, never surfacing as properties on the caller's generated classes, so the typed API stays free of hook internals. Batched hooks (batched: true) carry over: the HookLoader collects the typed data objects instead of scalar tuples.

This is a breaking change — the @hook(input:) argument is removed; DirectiveProcessor raises a migration error if it is still used. Hooks must move their field list into a requires fragment.