Rework type model: concrete wrappers + parallel Kind-lattice dispatch

[maleadt]

Today, @objcwrapper Foo <: Bar emits an abstract Foo plus a concrete FooInstance. Anything stored as a Foo (containers, struct fields, inferred return types) carries an abstract type, which boxes in Vector{Foo}, breaks inference through property chains, and triggers dynamic dispatch in profiles (cf. vchuravy's profile).

The reason we do this is because ObjC's class hierarchy is concrete inheritance with state at every level. For example, NSMutableString is a concrete class with fields and methods, and NSString is also a concrete class with fields and methods that NSMutableString extends. Julia's type system forbids this: only abstract types can have subtypes, only leaves can be concrete. There is no Julia equivalent to "a concrete struct that inherits from another concrete struct."

With this PR, each @objcwrapper Foo <: Bar now emits two parallel definitions:

struct Foo <: Object                        # concrete leaf (storage)
  ptr::id{Foo}
end
abstract type FooKind <: BarKind end        # parallel lattice (dispatch)
classkind(::Type{Foo}) = FooKind

Crucially, the inheritance stays flat: every wrapper is <: Object directly so that Vector{NSString} is bits-packed and inference sees fixed types through property access.

This of course breaks dispatch on "non-leaf" classes, for which we introduce a parallel Kind lattice used by @objcmethod methods:

@objcmethod foo(obj::KindOf{Bar}) = ...

The macro KindOf{T} lowers to trait dispatch on Type{<:classkind(T)} plus a method that forwards from untyped Object inputs. Multiple sites on the same function compose naturally -- Julia's dispatch picks the most specific Kind at the call site -- and downstream wrappers automatically participate by virtue of SubKind <: ParentKind. When the static type is known at the call site (the common case), the entry-then-body chain folds at compile time, so the trait dispatch is often zero-cost.

Closes #13

[maleadt]

JuliaGPU/Metal.jl#776 seems to be working fine, so let's try this out.

[christiangnrd]

Small nit but but the “@objcmethod no longer…” phrasing in the error message seems strange since this is the first release where the macro exists

[maleadt]

Yeah that's LLMs writing annoying comments based on interim state... Thanks for reading over.
In any case, @objcmethod is gone after #67.