Add payload lifecycle hooks and GC integration

Payload storage is now inline bytes only. Add optional lifecycle hooks (init/deinit) and GC integration for payloads that store Python refs.

Idea: detect payload methods via @hasDecl on the payload type, e.g.:

const BoxState = struct {
    value: u64,
    pub fn init(self: *BoxState) void {}
    pub fn deinit(self: *BoxState) void {}
    pub fn traverse(self: *BoxState, visit: py.ffi.c.visitproc, arg: ?*anyopaque) c_int {}
    pub fn clear(self: *BoxState) void {}
};

Hook points:

• call payload init in tp_new after allocation
• call payload deinit in tp_dealloc before free
• if traverse/clear provided, integrate into GC traverse/clear

Note: only required for payloads that manage resources or hold PyObject* refs; POD payloads remain zero-overhead.

Start by tracing payload storage and the tp_new and tp_dealloc hook points described in the issue. Determine how optional init/deinit and traverse/clear methods can be detected and connected, then verify that resource-owning and Python-reference payloads receive lifecycle and GC handling while POD payloads retain zero-overhead behavior.