gc/thread: iterator / generator / %TypedArray% intrinsic towers are built once per PROCESS, not per realm #8002

Same family as #7988 / #8001. One shape, several cells, one fix.

Each of these is a once-per-process if PTR.load() == 0 { build… } gate
whose result is a raw address of an object in the building thread's arena —
while the tower it names is a realm intrinsic and every perry/thread agent has
its own realm.

attach_iterator_prototype (iterator_prototypes.rs:270-288) then calls
chain_to(obj_ptr, proto_ptr) on every newly allocated iterator on every
thread, using whichever thread built the tower first — so a foreign pointer is
stored into a live object in this agent's own heap, which dangles once the
building thread exits. Any for…of / spread / Iterator.from hits it, i.e. it
is reachable from about the most common operation in the language.

Two of these carry comments that are now false:

• typed_array.rs:416 — "Called from populate_global_this_builtins
  (single-threaded under the singleton CAS)". Every thread's first
  js_get_global_this runs populate_global_this_builtins independently, so
  that assumption stops holding the moment a second OS thread exists.
• object/mod.rs:267-279 records that GC: the two lazy intrinsic-tower builders need #7217's no-move window, and need a gate that can fail first #7251 already hit this for tests and
  fixed it with per_test_global! — which, per per_test_global.rs:60-62,
  "in a non-test build … expands to the plain static it replaced, byte for
  byte". The production defect was never addressed.

scripts/gc_runtime_root_holders.json:90-130 marks the iterator cells
covered_elsewhere; that verdict is about GC-scan coverage
(object::scan_object_cache_roots_mut, object/mod.rs:1141) and says nothing
about the per-realm hazard, so it does not rule this out.

Suggested acceptance

A perry/thread test in which agent B runs for (const x of [1,2,3]) and a
function*, and a perry-runtime unit test in the shape of
a_second_agents_prototype_addresses_are_its_own (#7994) asserting two live
agents get DIFFERENT tower addresses, with per-agent liveness asserted before
distinctness.

The obstacle usually cited ("Darwin has no local-exec TLS") is stale — see
#8001's closing note and crates/perry-runtime/src/tls_hot.rs.